# Oryvelon — full facts Oryvelon is a digital company builder and venture group. It builds and operates independent businesses across software, marketplaces, AI products, commerce, digital services and education — including MerchNivo, CastLyra, EduRelia, ZodiVela, KeşifAtlası, Noveniq, WeAreMedia and Sinem Keser Beauty Academy. Oryvelon provides shared technology, AI, security, analytics and operating standards while keeping each company’s data, business logic and customer relationship separate. ## FAQ Q: What is Oryvelon? A: Oryvelon is a digital company builder and venture group. It builds and operates independent digital businesses across software, marketplaces, AI products, commerce and digital services. Q: Which companies are part of Oryvelon? A: MerchNivo, CastLyra, EduRelia, ZodiVela, KeşifAtlası, Noveniq, WeAreMedia and Sinem Keser Beauty Academy. Q: Is Oryvelon an agency? A: No. Oryvelon builds and operates its own companies. Agency services — web, Shopify, advertising, SEO and creative — are provided by WeAreMedia, an Oryvelon company. Q: Who founded Oryvelon? A: Oryvelon was founded by Uğur Keser and Sinem Keser. Q: Do Oryvelon companies share customer data? A: No. Each company keeps its own users, database and privacy boundary. Oryvelon does not build a cross-product user profile. Q: How does Oryvelon use AI? A: AI is used to explain, summarise and assist. Business-critical facts such as stock levels, visa rules or access permissions come from verified systems, not from a language model. Q: How can I contact Oryvelon? A: Use the form on the contact page or write to hello@oryvelon.com. For services, contact WeAreMedia directly. ## MerchNivo (merchnivo.com) An AI e-commerce employee for Shopify stores — products, inventory, orders, SEO and day-to-day operations in one assistant. Problem: Independent Shopify merchants run their stores across dozens of screens: product data, stock levels, orders, SEO, customer questions and, increasingly, advertising data. Most of that work is repetitive, and most small teams do not have an operations analyst to watch it. What it does: - Reads a store's products, inventory, sales and orders and turns them into clear daily priorities. - Flags operational issues — low stock, weak product pages, SEO gaps, unusual order patterns. - Helps with customer-support drafts and product content, grounded in the store's own data. - Expands over time to advertising performance, so merchandising and marketing decisions sit in one view. For: Shopify merchants and e-commerce teams that want an always-on operator without adding headcount. Business model: Recurring SaaS subscription with usage-based components. Principles: - Store facts — stock, prices, orders — come from Shopify and the product database, never from a language model's guess. - Every store is its own tenant. Store and customer data is never used for any other Oryvelon product. - Noveniq, an Oryvelon commerce brand, serves as a real operating store for development and case studies. ## CastLyra (castlyra.com) A global marketplace connecting brands with verified models, actors, creators, photographers and production talent. Problem: Brands looking for models, UGC creators, photographers or stylists still rely on scattered agencies, DMs and spreadsheets. Talent, in turn, struggles to be discovered by serious businesses — and to know that a request is genuine. What it does: - Talent applies with a structured profile and goes through verification. - Businesses discover talent by role, location, style and availability. - AI-assisted matching suggests relevant talent for a brief. - Requests, messaging and booking happen inside the platform with clear access rules. For: Brands, agencies and production teams on one side; models, actors, UGC creators, influencers, photographers, videographers, make-up artists and stylists on the other. Business model: B2B subscriptions for businesses, premium plans for talent, and booking commission over time. Principles: - Talent controls what is visible and to whom; businesses see contact details only under clear access rules. - Consent is explicit and recorded. - Talent data is never merged with any other Oryvelon product. ## EduRelia (edurelia.com) An AI learning platform for students, parents, teachers and schools — built with strict privacy and tenant isolation by design. Problem: Students learn at different speeds, parents want visibility, and teachers have limited time for individual support. Most AI tools for learning were not designed for schools — or for the care that children's data requires. What it does: - Personal learning support for students. - Progress visibility for parents. - Classroom tools and insight for teachers. - Administration and licensing for schools. For: Students and families directly, and schools licensing the platform for 100, 500 or 1,000+ students. Business model: Recurring family subscriptions and annual school licences. Principles: - The strictest privacy standard in the group: child-data minimisation and school-level tenant isolation. - Access is decided by authorization rules and row-level security — never by an AI model. - Student data is never used in any consumer product or for marketing. ## ZodiVela (zodivela.com) Personal discovery through astrology, numerology, tarot and traditional readings — made for reflection and entertainment. Problem: Millions of people enjoy astrology, tarot and traditional readings as a way to reflect. Most apps are either generic or make claims they cannot support. What it does: - Astrology, numerology and tarot readings. - Traditional formats such as yıldızname and coffee-cup reading. - Personalised, well-designed readings delivered in a modern consumer app. - Premium one-off readings alongside a subscription. For: Consumers who enjoy astrology and personal-discovery content. Business model: Recurring subscriptions, credits and one-off premium readings. Principles: - Positioned as entertainment and personal reflection — never as a prediction of the future or as medical, legal or financial advice. - User profiles are never combined with any other Oryvelon product. ## KeşifAtlası (kesifatlasi.com) Visa and relocation eligibility, grounded in a verified, versioned rules database — with clear personal reports. Problem: People planning to study, work or live abroad face scattered and outdated information. Generic AI chat answers about visas can be confidently wrong. What it does: - A free eligibility test. - A rules engine that evaluates answers against verified, versioned visa and residence rules. - A paid personal report delivered on the web, as a PDF and by email. - Connections to relocation partners where relevant. For: Individuals and families exploring study, work or relocation options — starting from Türkiye, built to go global. Business model: Paid personal reports, premium services and partner referrals. Principles: - Visa requirements come from a verified, versioned rules database — not from an AI model's memory. AI only explains the result. - Assessment data is minimised and never flows automatically into marketing profiles of other products. ## Noveniq (noveniq.com) A technology accessories brand, operated as an independent direct-to-consumer business. Problem: Great accessories are easy to find but hard to trust. Noveniq focuses on a curated range sold directly to customers. What it does: - A curated range of technology accessories. - Direct-to-consumer online store. - A real operating environment where Oryvelon tests commerce tooling such as MerchNivo. For: Consumers shopping for technology accessories online. Business model: Product sales margin. Principles: - Noveniq runs as an independent production business with its own customers and data. - Its role as a MerchNivo test and case-study store never changes how customer data is handled. ## WeAreMedia (wearemedia.us) A digital and advertising agency — web, Shopify, paid media, SEO, creative, content and AI automation for brands. Problem: Brands need one accountable partner across web, commerce and performance marketing. What it does: - Web and Shopify development. - Social media management, content and creative. - Meta Ads, Google Ads and SEO. - Branding, analytics and AI automation. For: Brands and e-commerce businesses that need hands-on digital execution. Business model: Service fees and retainers. Principles: - Oryvelon does not sell agency services directly — client work belongs to WeAreMedia. - Client data stays with each client engagement. ## Sinem Keser Beauty Academy (sinemkeser.com) A beauty and education brand — professional beauty services, appointments, training and digital learning products. Problem: People looking for professional beauty services want a trusted name and an easy way to book. Those who want to learn the craft need structured, practical education from an experienced professional. What it does: - Professional beauty services with online appointments. - Beauty education and training programmes. - Digital products for learners. For: Clients looking for professional beauty services, and students and professionals who want to learn beauty skills. Business model: Service fees, education programmes and digital product sales. Principles: - Client and student data stays with Sinem Keser Beauty Academy and is never shared with other Oryvelon companies. - Operates as an independent brand with its own website and customer relationship. # Insights (full text) ## Introducing Oryvelon: the group behind our companies URL: https://oryvelon.com/insights/introducing-oryvelon/ Category: Announcements Every group of companies eventually needs a name for the thing that holds it together. For us that name is Oryvelon. For years our work looked like a collection of separate projects. An online store here, an advertising operation there, a new product idea on a whiteboard, a domain bought late at night because the name was too good to lose. Each of those projects taught us something. Each of them also started more or less from zero: new hosting, new analytics, new email setup, new security decisions, new ways of handling data, new invoices from new tools. At some point the pattern became impossible to ignore. The hard, unglamorous parts of launching a digital business were the same every time. The parts that were genuinely different — the customer, the problem, the product, the brand — were the parts that deserved our attention, and they were getting less of it than they should. Oryvelon is our answer to that. It is the company builder that sits behind a small number of independent businesses. We build them, we own them and we operate them. And we do it on one shared set of foundations so that the next company starts further ahead than the last. What Oryvelon is The shortest description is the one on our homepage: we build and operate digital businesses. A slightly longer one: Oryvelon is a digital company builder and venture group. It starts companies from a validated problem, gives them the technology and operating standards they need, stays involved after launch and puts more resources behind the companies that are working. It helps to say what Oryvelon is not, because the alternatives are common: It is not an agency. We do not sell websites, social media management or advertising to clients from oryvelon.com. That work belongs to WeAreMedia, our services company, and anyone looking for those services goes there. It is not a single SaaS product. There is no "Oryvelon app" that tries to do everything. Each company has its own product. It is not a portfolio showcase. The Companies page lists real brands we operate, not ideas we might build one day. It is not a passive holding structure. We are involved in how each company is built, measured, secured and funded. If you want the longer comparison between these models, we wrote it up in Company builder, venture studio or holding company?. The companies Today the group includes eight companies, each with its own brand and website: Company What it does Category MerchNivo An AI e-commerce employee for Shopify stores Software CastLyra A global marketplace for verified creative talent Marketplace EduRelia AI learning for students, families and schools AI product ZodiVela Personal discovery through astrology, tarot and traditional readings Consumer AI KeşifAtlası Visa and relocation eligibility grounded in verified rules AI product Noveniq A technology accessories brand sold direct to consumers Commerce WeAreMedia A digital and advertising agency for brands Digital services Sinem Keser Beauty Academy Beauty services, education and digital products Beauty & education They look very different from one another, and that is intentional. A Shopify merchant, a school administrator, a model looking for work, a family researching a move abroad and someone booking a beauty appointment have almost nothing in common as customers. What they share is that behind each product there is a team that has thought carefully about infrastructure, data, security and cost — because that team is Oryvelon. Why a company builder, and why now There are three reasons we chose this structure instead of running each business on its own. The foundations are the same every time Almost every digital business needs the same layer underneath it: domains and DNS, SSL, deployment, authentication, email delivery, analytics, error monitoring, backups, access management, AI model access and cost tracking. None of that is what customers pay for. All of it can go wrong in expensive ways. When you run one company, you build that layer once and live with its quirks. When you run several, you either rebuild it every time or you build it properly once and reuse it. Oryvelon exists so that we can do the second. AI changed the economics of small teams Modern AI models make it possible for a small team to build products that would have needed a much larger organisation a few years ago. MerchNivo can act as an operations assistant for a Shopify store. KeşifAtlası can explain complex visa rules in plain language. EduRelia can support learning at the pace of each student. The same technology also creates new risks: models that confidently invent facts, costs that grow silently, prompts that change behaviour without anyone noticing, data that ends up in the wrong place. Handling those risks well is a discipline in itself. It makes sense to build that discipline once, at the group level, and apply it to every product. We describe our approach in AI explains, verified data decides. Independence has to be designed It is easy to say that each company is independent. It is harder to make it true. If all your products share one database, one login and one codebase, they are not really independent — they are features of one app with different logos. We want every company in the group to be able to stand on its own. One day a company might raise capital, bring in a partner or be sold. That only works if its code, data and infrastructure can be separated cleanly. So we design for independence from the first day, as we explain in Why we don't merge our products into one app. How Oryvelon works in practice Our operating model has six stages: Discover, Validate, Build, Launch, Operate, Scale. They are simple to list and demanding to follow. Discover means starting from a real problem and a specific customer, never from a trend. Validate means proving demand and a revenue mechanism before serious engineering begins. Build means shipping a first version on our shared foundations with clear acceptance criteria. Launch means going live under the company's own brand, domain and customer relationship. Operate means running the company on real metrics, with its own profit and loss and a 30/60/90-day review rhythm. Scale means putting capital and people behind what works — and stopping what does not. Before any idea becomes a company, it passes a ten-step protocol: problem and target user, revenue mechanism, recurring income loop, brand and domain ownership, a written source of truth, naming standards, data sensitivity classification, a phase-zero website, MVP acceptance criteria and 30/60/90-day KPIs. We published the full list in How we start a new company. What the group shares — and what it never shares This is the heart of the model, so it is worth being precise. Shared across companies: Domain, DNS, SSL and edge security standards Code review, repository and deployment standards An AI gateway for model routing, budgets, logging and fallbacks Error monitoring and product analytics tooling Brand-neutral interface building blocks and security headers Access management, secrets handling and cost reviews Kept separate, always: Each product's user database Customer, store and student data Prompts and private knowledge bases Business rules and brand voice API keys, budgets and logs Marketing audiences The short version is that we share the plumbing, not the people. A customer of one company is not quietly turned into a profile in another. We explain why in Shared infrastructure, separate data. The principles we hold ourselves to A group like ours could easily drift into bad habits: launching too many things, blending data because it is convenient, letting AI make decisions it should not make, or hiding weak products inside a combined bill. These principles are there to stop that. Real problems, real revenue. Every company needs a clear customer, a revenue engine and measurable product metrics with its own P&L. Independent companies. Each company could stand alone, and our architecture makes that realistic. Shared foundations, separate data. Infrastructure is shared. Data, keys, prompts and privacy boundaries are not. AI explains, verified data decides. Stock levels come from the store, visa rules from a verified rules database, access rights from authorization — never from a language model's guess. Visible economics. Every product sees its own hosting, database, AI, email and tooling costs next to its revenue. A group invoice must never hide whether a product is healthy. Only real companies on this website. Ideas stay internal until they are real brands. Who is behind Oryvelon Oryvelon was founded by Uğur Keser and Sinem Keser. Uğur's background is e-commerce, Shopify, AI and automation — many years of building stores, running advertising operations and turning manual work into systems. You can read more on ugurkeser.art. Sinem's background is brand, social media, creators and fashion, along with beauty and education through Sinem Keser Beauty Academy. More on sinemkeser.com.tr. Between them, the two backgrounds cover the two halves every digital business needs: the systems that make it run and the brand that makes people care. What the operating core actually does "Shared foundations" can sound abstract, so here is what the operating core does on an ordinary week. It keeps domains renewed, locked and pointed at the right place. It makes sure every sending domain has the email authentication records that stop other people from sending mail in our name. It maintains the deployment standards that let a company ship a change, preview it and roll it back if something breaks. It watches error reports and uptime. It reviews who has access to which system, and removes access that is no longer needed. It tracks what each company spends on hosting, databases, AI models, email and third-party tools, and puts that next to what the company earns. On the AI side, it decides which model families each product may use, sets budgets per product, keeps prompts versioned with an owner and a change log, and tests important changes against fixed evaluation sets before they reach users. When a model fails or an output looks risky, there is a defined fallback or a human path. None of this is visible to a customer of ZodiVela or CastLyra. All of it is the reason those products can be trusted. On the growth side, the core makes sure every company is discoverable in the right way: clean site structure, structured data that tells search engines and AI assistants what a company is, sitemaps, sensible robots rules, and measurement that respects visitors. We cover that in more detail in Search and AI-assistant visibility. How we will know if the model works A company builder is a bet that shared foundations make each new company better and cheaper to run than it would be alone. We intend to hold ourselves to that bet with plain questions rather than slogans. Does each new company launch with fewer surprises than the last? Security gaps, missing email records, broken analytics and forgotten renewals should become rarer, not more common. Can we see every company's economics on its own? If we cannot say whether a product is healthy without digging through a combined bill, the model is failing. Do the boundaries hold? A customer of one company should never discover that their data turned up in another. That is not a metric we are willing to trade off. Are we stopping things? A portfolio that never closes or reshapes anything is not being managed. The continue, stop or scale decision has to be real. Is AI making products better without making them less reliable? Speed and convenience are welcome. Invented answers are not. Questions founders and partners ask us We are often asked a handful of the same questions, so here are short answers. Do you build companies for other people? No. Oryvelon builds and operates its own companies. If you need a website, a Shopify store or advertising, that is WeAreMedia's work. Why are the companies so different from each other? Because the shared layer is technical and operational, not a product. A marketplace and a school platform can share deployment standards, security reviews and AI governance without sharing anything a customer would see. Will the companies eventually merge into one platform? No. That would weaken each of them. We explain the reasoning in Why we don't merge our products into one app. How do you choose what to build next? Every idea goes through the same protocol. Most ideas stop early, which is exactly what the protocol is for. What you will find on this website Oryvelon.com is deliberately simple. It explains the group, lists the companies and shares how we work. Companies — every Oryvelon company with a short profile, its category, business model and data principles. About — the company builder model and our principles. Build — the operating model and new-company protocol. Insights — long-form notes on company building, AI operations, data and security, commerce, marketplaces and growth. Contact — for partnerships, media and investment conversations. We write Insights to be useful on their own. They explain how we think about problems that many founders and operators face: how to validate an idea, how to run AI in production without losing control of cost or accuracy, how to keep customer data where it belongs, how to measure traffic between sister brands without building a surveillance machine. Where we use our own companies as examples, we describe what they do and how they are designed. Three ordinary situations, and how the group handles them Principles are easy to write. Here is what they look like when something actually happens. These are illustrative situations of the kind we plan for, not reports of specific incidents. A new stock signal for MerchNivo. Suppose the MerchNivo team wants to add a warning when a best-selling product is about to run out. The first place it gets tried is Noveniq, the technology accessories store we operate ourselves, because a real store with real orders shows problems a demo never will. Noveniq is still treated as a separate tenant: its data trains nothing else and flows nowhere else. When the signal behaves well there, it ships to MerchNivo customers. We explain the approach in Using a real store as a product test-bed. A visa rule changes. A requirement that KeşifAtlası relies on is updated by the relevant authority. The fix is not a new prompt. Someone updates the rule in the verified rules database, records the source and the effective date, and the previous version stays on file. The AI layer keeps explaining whatever the rules say. See Rules engines for eligibility decisions. Someone asks Oryvelon to build their website. The answer is polite and short: that is WeAreMedia's work, and we pass the request there. Keeping service work inside one company keeps its economics honest and keeps the group's role clear. We describe the arrangement in Running a services company inside a company builder. What every new company inherits on day one A company that joins the group does not start with a blank page. It starts with: a registered, locked domain and email authentication records; a written source-of-truth document that later decisions must agree with — see Why every company starts with a source-of-truth document; a phase-zero website that states what the company does before the product is finished, as described in Phase-zero sites; deployment, monitoring and access standards, with two-factor authentication from the first account; its own AI project, keys and budget if it uses models; its own cost lines, so its economics are visible from the first invoice. What it does not inherit is anyone's customers. That part it has to earn. What comes next The companies will keep evolving. Some will grow quickly and receive more of the group's attention. Some may change shape. New ones will only appear on this site after they have passed the protocol and become real brands. The foundations will keep improving too. Every lesson one company learns — about security, deployment, analytics, AI cost or data handling — becomes a standard the next company inherits. That compounding effect is the reason Oryvelon exists. If you are interested in partnering with the group or one of its companies, writing about what we do, or talking about investment, we would like to hear from you. The contact page is the fastest way to reach us. If you need web, Shopify, advertising or SEO services, WeAreMedia is the right company to talk to. Thank you for reading. We are glad you are here. ## What is a digital company builder? URL: https://oryvelon.com/insights/what-is-a-digital-company-builder/ Category: Company Building When people hear that Oryvelon is a "company builder", the follow-up question is almost always the same: so what exactly do you do all day? It is a fair question. The term is used loosely, sometimes for agencies, sometimes for investment firms, sometimes for groups that are really one product with several logos. This note is our attempt to define the term clearly, explain why the model exists, show where it goes wrong, and describe how we apply it at Oryvelon. A working definition A digital company builder is an organisation that repeatedly starts, owns and operates digital businesses, using a shared set of people, technology and operating standards so that each new business starts further ahead than the last. Every part of that sentence matters. Repeatedly starts. One successful company does not make a company builder. The model only makes sense if you intend to build more than one, and if you have a repeatable way of doing it. Owns. The companies are part of the group. They are not client projects and they are not minority investments in someone else's idea. Operates. The builder stays involved after launch. It watches metrics, maintains infrastructure, reviews security, controls cost and decides where to put more resources. Shared foundations. The advantage comes from reusing what is common: tooling, standards, experience and people who have solved the same problems before. Starts further ahead. If the fifth company is no easier to launch than the first, the model is not working. Where the model comes from Groups that build several businesses under one roof are not new. Industrial conglomerates, media groups and family holding companies have done versions of it for a long time. What changed with digital businesses is the cost structure. A software product, a marketplace or an online store has a large fixed layer that is surprisingly similar from one business to the next: hosting, domains, deployment, authentication, email, analytics, monitoring, payments, legal pages, security, backups. Twenty years ago much of that had to be built by hand. Today most of it can be assembled from reliable services — but assembling it well still takes experience, and assembling it badly creates problems that surface months later. A company builder takes advantage of that. It learns how to assemble the foundation properly, writes it down as a standard and reuses it. The business-specific work — understanding the customer, designing the product, building the brand — gets more attention because the foundation no longer eats the first months of every project. The arrival of capable AI models has strengthened the case further. Small teams can now build products that once required large organisations. But AI brings its own operational discipline: model choice, cost control, prompt management, evaluation, safety and data boundaries. Building that discipline once and applying it across several products is exactly the kind of leverage a company builder is designed for. How it differs from similar models The easiest way to understand a company builder is to compare it with the models it is often confused with. Model Main activity How it earns Key difference Agency Delivers services for clients Fees and retainers Revenue comes from client work, not owned products Holding company Owns stakes in businesses Dividends and exits Often passive; does not build or operate the businesses Venture studio Creates startups, often with external founders Equity in spun-out companies Hands over control earlier Venture capital fund Invests in companies built by others Returns on investment Does not build anything itself Single SaaS company Runs one product Product revenue One customer base, one product, one set of economics Company builder Builds, owns and operates several companies Value of the companies it runs Keeps operating control and shares foundations We wrote a longer piece on these distinctions in Company builder, venture studio or holding company?. The short version: an agency sells its time, a fund sells its capital, a holding company owns, a studio launches, and a company builder does all of the building and keeps doing the operating. At Oryvelon we are strict about the agency distinction in particular, because it is the one most likely to cause confusion. Oryvelon does not sell services to clients. The group has a services company — WeAreMedia — and that is where client work lives. What a company builder actually shares "Shared foundations" is only useful if it is specific. Here is what a well-run company builder typically shares across its companies, and how it looks at Oryvelon. Infrastructure and security standards Domains are registered with renewal and locking in place. DNS follows one standard. Every site is served over HTTPS with sensible security headers. Every sending domain has SPF, DKIM and DMARC so nobody can easily send email in its name. Access to critical systems uses two-factor authentication and least privilege. These sound like small things. Together they remove a whole category of avoidable incidents. We describe the domain side in Domain, DNS and email security for a multi-brand group. Engineering and deployment practice Each company has its own repositories and release cycle, but they follow the same conventions: preview environments, staging where it matters, production changes that can be rolled back, and a clear rule that production data is not copied into test environments. See Local, preview, staging, production. AI access and governance Every AI feature in the group goes through the same kind of gateway: model routing, budgets, logging, fallbacks and schema validation. But — and this is critical — each product keeps its own prompts, knowledge bases, keys and logs. We explain the design in What an AI gateway does — and what it should never do. Measurement Each company can see its traffic, conversions and costs on its own. Traffic between sister brands is measured with consistent UTM parameters so we understand how the group helps each company, without forcing users into a single identity graph. See UTM standards for measuring traffic between sister brands. Operating rhythm New companies go through the same protocol, and running companies go through the same review cadence. That consistency makes it possible to compare companies fairly and decide where resources should go. See The operating cadence behind a portfolio of digital businesses. What a company builder should never share The opposite list matters just as much. A company builder that shares the wrong things loses the trust that makes each company valuable. Customer data. A person who buys from one company has not agreed to become a profile in another. User databases. Separate products should not live in one database table with a "brand" column. Prompts and private knowledge. A consumer app's prompt has no business being reused in a school platform. Business rules. Eligibility rules, pricing logic and access policies are part of each company's core, not shared helpers. Brand voice and audiences. Each company speaks to its own customers in its own way. At Oryvelon we summarise this as: share the plumbing, not the people. It is written into our operating standards and explained in Shared infrastructure, separate data and Tenant isolation explained. Why the model works When it is done well, a company builder produces three advantages that a standalone startup struggles to match. Speed without shortcuts. Because the foundation already exists, a new company can reach a working phase-zero site and first product version quickly — and without skipping the security, privacy and measurement steps that rushed teams usually skip. Better decisions about capital. A group that can see the real economics of each company can move money and attention towards the ones that are working. A standalone founder has one company and therefore no comparison. See Capital allocation inside a company builder. Compounding knowledge. Every incident, every good vendor choice, every lesson about pricing or onboarding becomes a standard for the next company. Over time the group gets better at building companies in general, not just at running one. Where the model goes wrong We think it is important to be honest about the failure modes, because we designed Oryvelon specifically to avoid them. Forcing everything into one app It is tempting to give every product the same login, the same database and the same codebase. It feels efficient. It is usually a mistake. Different users, different promises and different release cycles pull a shared app apart, and the result is a platform nobody loves and no company can leave. We wrote about this in Why we don't merge our products into one app. Merging customer data The second temptation is to combine what you know about people across products. It might look like a marketing advantage. In practice it erodes trust, creates regulatory risk and makes it much harder to ever sell or partner a single company. Spreading budget evenly A company builder that gives every idea the same budget is not allocating capital; it is avoiding decisions. Resources should follow evidence. That means some companies get more, and some ideas are stopped. Launching too many things Because a builder can launch quickly, it can also launch too much. A portfolio of half-finished companies is worse than a few that are genuinely run. Our rule is simple: ideas stay internal until they pass the protocol, and only real brands appear on the Companies page. Letting the core become a bureaucracy The shared layer is there to help companies move faster. If every change needs permission from a central team, the model has turned into overhead. Standards should be written down, easy to follow and revisited when they stop being useful. How Oryvelon applies the model Here is how the model looks in our case. The companies. Oryvelon builds and operates MerchNivo, an AI e-commerce employee for Shopify stores; CastLyra, a global marketplace for verified creative talent; EduRelia, an AI learning platform for students, families and schools; ZodiVela, a consumer app for astrology, tarot and traditional readings; KeşifAtlası, a visa and relocation eligibility platform grounded in verified rules; Noveniq, a technology accessories brand; WeAreMedia, a digital and advertising agency; and Sinem Keser Beauty Academy, a beauty and education brand. The operating model. Every company moves through Discover, Validate, Build, Launch, Operate and Scale. Before an idea becomes a company it passes a ten-step protocol, published in How we start a new company. The core. Oryvelon provides the shared standards for domains, DNS, security, deployment, AI access, analytics and cost control. It does not run each company's day-to-day customer operations. The boundaries. Each company has its own database, keys, prompts, knowledge and customer relationship. There is no cross-product user graph. The economics. Each company's costs and revenue are tracked separately so the group can see which companies are healthy and which need attention. A worked example: what the builder does, step by step Definitions are easier to understand with an example. Here is a simplified walk-through of how a new company moves through Oryvelon, using the kind of decisions we actually make. 1. A problem shows up repeatedly. Through years of running e-commerce operations, the same pain kept appearing: store owners drowning in product data, stock levels, orders, SEO tasks and support questions, without anyone watching the whole picture. That is a real problem with a specific customer — the independent Shopify merchant — and it became the starting point for MerchNivo. 2. The idea is written down before it is built. The first artefact is not code. It is a one-sentence problem statement, a description of who pays and why, and a sketch of how revenue repeats — in MerchNivo's case, a monthly subscription with usage-based components. If we cannot write those three things clearly, the idea is not ready. 3. The brand and domain are secured. A company that cannot own its name cannot build trust. Domains are registered, locked and set to renew. The brand is checked against existing companies in its category. 4. The data sensitivity is classified. A store-operations product will hold store data, product data and customer order data. That means store-by-store tenant isolation from the first version. Deciding this before any code is written is much cheaper than retrofitting it later — see Tenant isolation explained. 5. A phase-zero site goes live. Before the product is finished, the company gets a real website with real content, a way to join a waitlist and proper search visibility. This tests whether the message resonates and starts building an audience. See Phase-zero sites. 6. The first version is built on the shared foundation. Deployment, monitoring, email, analytics and AI access come from the group standard. The team spends its time on what makes the product different: connecting to the store, understanding the data, turning it into useful priorities. 7. The company is measured on its own. From launch, the company has its own metrics and its own cost lines. At 30, 60 and 90 days we look at the evidence and decide whether to continue, stop or scale. 8. Lessons flow back to the core. If MerchNivo discovers a better way to handle a class of AI errors, or a cleaner approach to background jobs, that becomes part of the standard the next company inherits. The same sequence applies to very different companies. For KeşifAtlası, the critical early decision was that visa rules must come from a verified, versioned rules database rather than from an AI model. For EduRelia, it was that student data requires the strictest privacy standard in the group. For CastLyra, it was that verification and consent had to be designed into the marketplace from the start. The builder's job is to make sure each of those decisions is made early, deliberately and in writing. What people who work with a company builder notice Partners and collaborators tend to notice three things when they work with a well-run company builder. First, decisions are written down. There is a source-of-truth document for each company, and later documents are expected to agree with it. That makes it easier to join a project mid-way and understand why things are the way they are. We explain the practice in Why every company starts with a source-of-truth document. Second, the boring parts are already solved. Nobody spends the first week of a collaboration arguing about which analytics tool to use, how to name repositories or where secrets are stored. Those answers exist. Third, the conversation is about the customer. Because the foundation is settled, most discussions focus on the problem, the product and the brand — which is where they should be. Is a company builder right for every founder? No. If you have one idea you care deeply about and want to spend the next ten years on it, a single focused company is probably the right structure. A company builder makes sense when three conditions are true: You expect to build more than one business over time. The businesses share a meaningful technical or operational foundation. You are willing to make hard allocation decisions between them. We meet those conditions. Our backgrounds span e-commerce, advertising, brand, creators, education and AI, and many of the problems we want to solve share the same digital foundation. That is why Oryvelon exists in the form it does. A simple test for any company builder If you are evaluating a company builder — as a partner, an employee, an investor or a journalist — here are five questions that separate real ones from labels: Which companies do you own and operate today? A real builder can name them and show them. What exactly is shared? The answer should be specific and technical, not "synergies". What is never shared? If the answer does not include customer data, be careful. How do you decide to stop a company? A builder that has never stopped anything is not allocating capital. Can each company stand alone? If the answer is no, you are looking at one product with several brands. We try to answer all five openly on this website. The Companies page lists what we run. The Build page explains how. This Insights section explains why. Three common misreadings of the term Because "company builder" is used loosely, three misreadings come up often. "It is an agency with equity." No. An agency's main output is work delivered to clients. A company builder's main output is companies it owns and keeps running. "The companies must share customers." They do not have to, and in our case they deliberately do not. The shared part sits below the customer. "Shared foundations mean shared risk." Only if the foundations are built badly. Separate data, separate keys and per-company budgets mean that a problem in one company is contained there, while a fix to the standard reaches all of them. Summary A digital company builder is not magic and it is not a buzzword. It is a practical way to build several businesses without rebuilding the same foundations every time — as long as it is disciplined about what it shares, what it protects and where it puts its resources. That discipline is what Oryvelon is built around. ## Why we don't merge our products into one app URL: https://oryvelon.com/insights/why-we-dont-merge-products-into-one-app/ Category: Company Building There is a moment in the life of every multi-product group when someone draws a diagram with one big box in the middle labelled "platform" and all the products hanging off it. One login. One database. One codebase. One design system. One customer profile. It looks tidy. It looks efficient. It looks like the grown-up thing to do. We have drawn that diagram too — and then deliberately decided not to build it. This note explains why Oryvelon keeps its companies as independent products, what we share instead, and how we decide where the line sits. If you are running more than one digital product, the reasoning may be useful to you as well. The appeal of the super app It is worth taking the idea seriously before arguing against it, because the appeal is real. A single platform promises one codebase to maintain instead of several. It promises one login, so a customer who uses two products does not need two accounts. It promises one customer profile, which marketing teams love because it suggests you can sell more to the same people. It promises one design system, so everything looks consistent. And it promises lower cost, because shared infrastructure is cheaper than duplicated infrastructure. Some of those promises are true. The cost argument in particular is strong — which is exactly why we share infrastructure. But the other promises tend to break down when the products serve genuinely different people. Different users, different promises Look at the people who use Oryvelon companies: A Shopify merchant using MerchNivo wants an operations assistant that reads their store data and helps them run the business. A school administrator using EduRelia is responsible for children's data and needs the strictest privacy guarantees we can offer. A photographer or model on CastLyra wants control over who sees their profile and contact details. Someone using ZodiVela is looking for reflection and entertainment through astrology or tarot. A family on KeşifAtlası is exploring whether they can study, work or live in another country. A customer of Noveniq is buying a phone accessory. Each of these products makes a different promise about what it does and what it does with personal information. A school cannot accept that its students' data sits in the same database as a consumer entertainment app, even if the tables are separate. A talent profile should not be one query away from a merchant's customer list. A person exploring relocation options should not find that their answers are being used to market horoscopes. In a single app, those promises blur. Even if the engineering is careful, the perception is damaged — and in our experience, perception is where trust lives or dies. Different lifecycles Products also move at different speeds. A consumer app like ZodiVela benefits from frequent experiments: new reading formats, new designs, new pricing tests. An education product like EduRelia has to be conservative. Schools plan terms in advance; teachers need stability; changes to how student data is handled require care and documentation. A marketplace like CastLyra has two sides to keep in balance, and changes that help businesses can hurt talent if they are not thought through. If all of these lived in one codebase with one release pipeline, the fastest product would be slowed down by the most careful one, or the most careful one would be exposed to risks created by the fastest. Neither outcome is acceptable. Separate repositories and separate release cycles let each company move at the speed that suits its customers. The shared part — how we deploy, test, preview and roll back — is a standard, not a single pipeline. We describe that standard in Local, preview, staging, production. A company must be able to stand alone This is the reason that matters most to us as owners. Any company in the group could one day raise its own capital, bring in a strategic partner or be sold. We are not planning any of that today, but we want it to be possible. It is only realistic if the company's code, data and infrastructure can be separated cleanly. Think about what separating a product from a super app involves. You would need to extract its users from a shared identity system, split its records out of shared database tables, untangle its code from shared modules full of product-specific exceptions, migrate its analytics history and rebuild its deployment. It is the kind of project that takes months, costs a lot and rarely goes perfectly. Now think about separating a product that was independent from the start. Its repository already belongs to it. Its database already contains only its data. Its keys, budgets and logs are already its own. Separation becomes an administrative exercise rather than a surgical one. We go into this in more depth in Designing every company so it could stand alone. What we share instead Independence does not mean everyone rebuilds everything. That would throw away the main advantage of a company builder. The trick is to share the right layer. Brand-neutral building blocks We share basic interface primitives and design tokens — spacing, typography scales, form components, accessibility patterns. Each company applies its own brand on top. A MerchNivo dashboard and a ZodiVela reading screen should not look alike, but both should have accessible forms and sensible spacing. Authentication wrappers, not user databases We share code that makes it easy to add secure login to a product. We do not share the user table. Each product has its own users, and a person who uses two Oryvelon products has two separate accounts by default. Logging, tracing and validation utilities Every product needs to log errors, trace slow requests and validate input. The utilities for that are shared. What they log stays with each product. Security headers and deployment standards The HTTP headers that protect a site, the rules for secrets, the standards for DNS and email authentication, the way environments are structured — all shared. See Secrets management and Domain, DNS and email security. An AI gateway client Every AI feature goes through a common gateway that handles model routing, budgets, logging and fallbacks. Each product uses it with its own project, keys, prompts and limits. The gateway standardises how we call models; it does not merge what we send to them. See What an AI gateway does. Email delivery utilities Transactional email — receipts, login codes, notifications — is sent through the same kind of service with the same authentication standards, but from each company's own domain. What we never share The list of things we keep separate is just as deliberate: Never shared Why Product user databases Each company owns its customer relationship Customer, store or student data Consent is given to one company, not the group Prompts and private knowledge bases They encode each product's expertise and tone Business rules Pricing, eligibility and access logic are each company's core Brand voice Customers should feel they are dealing with a distinct brand Marketing audiences Using one company's customers to market another breaks trust API keys, budgets and logs Each product's costs and incidents must be visible on their own We sometimes summarise this as: share the plumbing, not the people. "But wouldn't a single login be more convenient?" This is the most common objection, and it deserves a direct answer. For some groups, a single account across products is a real benefit — think of a suite of productivity tools used by the same team. For us, the overlap between customers is small and the privacy expectations are very different. The convenience of a shared login would be felt by few people; the loss of clarity about who holds your data would be felt by many. If a genuine need for linked accounts appears between two specific companies, it can be designed as an explicit, consented connection between two independent systems. That is very different from building a single identity layer that every product depends on. "And what about cross-selling?" The second objection is commercial: surely a group should sell more to the customers it already has? We do believe in helping people discover the group's other companies — in the right context. The Oryvelon website links to every company. A founder's personal site can point to relevant products. A company's footer can say "An Oryvelon company". We measure that traffic carefully with consistent UTM parameters, as described in UTM standards for measuring traffic between sister brands. What we do not do is take a customer's data from one company and use it to target them for another. Marketing consent is not product consent, and treating them as the same thing is a short-term gain with a long-term cost. How we decide where the line sits New situations come up regularly: a helper library that might be shared, a feature that two companies both want, a vendor that could serve the whole group. We use three questions to decide. Is it brand-neutral? If the component carries a product's look, voice or business logic, it belongs to that product. Is it privacy-safe? If sharing it means data from one product could reach another, it stays separate. Does it preserve independent deployment? If a change in the shared component could force every product to release at once, it needs a looser coupling — versioning, or simply duplication. If the answer to all three is yes, it can be shared. If any answer is no, it stays with the company. Four decisions, worked through Principles are easier to trust when you can see them applied. Here are four realistic decisions and how the three questions resolve them. A shared date-picker component. Several products need a calendar input: CastLyra for booking dates, Sinem Keser Beauty Academy for appointments, EduRelia for class schedules. The component is brand-neutral once styling is applied per product, it holds no data, and it can be versioned so each product upgrades when it is ready. Verdict: share. A shared "customer insights" service. Someone suggests a central service that collects behaviour from every product to build richer profiles. It is not privacy-safe by design — its whole purpose is to combine data across companies. Verdict: do not build. Each product keeps its own analytics, and the group looks at aggregated, product-level metrics instead. A shared prompt for summarising text. Two products both need to summarise user-provided text. The prompt looks similar, but each product's tone, safety requirements and output format differ: a store summary for a merchant is not a reading summary for a consumer. The technique can be shared as documentation; the prompt itself belongs to each product and is versioned there. Verdict: share the pattern, not the prompt. See Prompt versioning and evaluation. A shared email-sending utility. Every product sends transactional email. A small library that formats messages, handles retries and logs delivery status is brand-neutral and holds no data beyond what each product passes to it at send time. Each product still sends from its own domain with its own authentication records. Verdict: share. Notice the pattern. Code that helps each company do its own work tends to be shareable. Anything whose purpose is to combine what companies know about people is not. A checklist for teams running more than one product If you run several products and are tempted by the super-app diagram, these questions can help you decide. Do the products serve the same people in the same role? If not, a shared identity layer will mostly add risk. Do they make the same promise about data? If one handles children's data or sensitive personal circumstances, it needs its own boundary. Do they need to release at the same pace? If not, a shared pipeline will slow one down or expose another. Could you ever want to sell, spin out or partner one of them? If yes, independence is cheaper to build now than to extract later. Can you see each product's costs and revenue separately today? If not, fix that before any consolidation, or you will never know whether consolidation helped. Is the thing you want to share plumbing or product? Plumbing is usually safe to share. Product rarely is. Most groups that answer these honestly end up somewhere in the middle: shared standards and libraries, separate products and data. That is where Oryvelon sits. The cost of saying no to the super app Independence is not free. There is some duplication. Two products may each implement a similar settings screen. Each has its own database to back up and monitor. Each has its own analytics setup. Occasionally a bug fixed in one product has to be fixed separately in another. We accept that cost because the alternative costs more in the ways that matter: trust, speed and optionality. And we reduce it where we can — shared libraries, shared standards, shared monitoring practices and a small group of people who know how all of it fits together. It also helps that we track each product's economics on its own. When costs are visible per company, it is easy to see whether independence is becoming expensive anywhere, and to address it specifically rather than by merging everything. See Unit economics per product in a shared-infrastructure group. What this looks like from the outside If you use two Oryvelon products, you will notice: two separate brands, with their own websites and domains; two separate accounts, if the products need accounts; two separate privacy notices, each describing that company's data practices; and, at the bottom of each site, a small note saying it is part of Oryvelon. That last detail is intentional. We want people to know the companies are connected, because the group stands behind each of them. We just do not want that connection to come at the expense of their privacy or of each company's identity. Mistakes that quietly undo independence Deciding not to build a super app is the easy part. Most groups that lose independence do it gradually, through small shortcuts that each look harmless. These are the ones we watch for. One admin account for every product. A single login that can reach every company's hosting, database and payment dashboard is a shared identity layer in all but name. Each company gets its own accounts, with two-factor authentication and the narrowest role that does the job. See Least-privilege access for small teams. Copying code without versioning it. When a helper is pasted into three repositories, the copies drift. When it is imported unpinned, a change in one place breaks three products at once. Shared libraries are versioned, and each company upgrades on its own schedule. A "temporary" shared database. Two early products put in one database "until we split it later" rarely get split. Separation is cheapest on the first day. We explain the options in Tenant isolation explained. One billing account hiding everything. If hosting, AI and email costs for every company land on one invoice with no breakdown, nobody can tell which product is healthy. Support tooling that sees everything. A shared help desk inbox where one agent can open any company's customer records turns separate products into one from the customer's point of view. Consider how this plays out with Noveniq and MerchNivo. The two work closely, because Noveniq is where MerchNivo features meet a real store first. It would be convenient to give the MerchNivo team direct access to Noveniq's store admin. Instead, Noveniq connects to MerchNivo the way any merchant would, as a tenant with its own permissions. The friction is small, and it keeps both companies honest about what a real customer experiences. Summary We share infrastructure because it makes every company stronger and cheaper to run. We do not merge products because it would make each of them weaker, slower and harder to trust. The line between the two is not always obvious, so we draw it deliberately, write it down and revisit it when the situation changes. The companies are listed on the Companies page. The operating model that keeps them independent is described on the Build page. ## AI explains, verified data decides URL: https://oryvelon.com/insights/ai-explains-verified-data-decides/ Category: AI Operations Language models have a strange combination of strengths. They can read a messy product catalogue and describe it clearly. They can take a dense visa regulation and explain it to a family in plain language. They can turn a teacher's notes into a friendly progress summary for a parent. They can draft, rephrase, translate and structure faster than any person. They can also state something false with complete confidence. That combination shapes every AI feature we build at Oryvelon. Our rule is short enough to fit on a sticky note: AI explains, verified data decides. This note explains what the rule means, why we hold to it, and the engineering practices that make it real rather than aspirational. The problem in one example Imagine a Shopify merchant asks an AI assistant: "How many units of the black phone case do I have left?" A language model on its own has no idea. It has never seen the store. But it has seen millions of sentences that look like answers to questions like this, so if you let it, it will produce one: "You have 42 units remaining." The number is fluent, plausible and invented. For a chat demo, that might be amusing. For a merchant deciding whether to reorder stock, it is a real cost. For a family deciding whether to apply for a visa, or a school deciding who may see a child's work, it is unacceptable. The fix is not to make the model "more careful". The fix is to never ask the model for the fact in the first place. What "verified data decides" means For each product, we identify the facts that are business-critical — the ones where a wrong answer causes real harm — and we make sure those facts come from a system designed to hold them. Product Business-critical fact Where it comes from MerchNivo Stock levels, prices, orders The Shopify API and the store database KeşifAtlası Whether a visa or residence requirement applies A verified, versioned rules database EduRelia Who may see a student's work Authorization rules and row-level security CastLyra Whether a business may see a talent's contact details Consent records and access rules ZodiVela What a user has purchased or has credits for The billing system In each case, the model can talk about the fact. It can explain it, put it in context, suggest what to do next. It is never asked to produce it. What "AI explains" means Taking facts away from the model does not make it less useful. It makes it useful in the right way. In MerchNivo, the store data says three products are low on stock and one product page has no description. The model turns that into a clear morning briefing: which items to reorder first, why the missing description matters for search, a suggested draft for that description based on the product's attributes. In KeşifAtlası, the rules engine determines which routes a person may be eligible for and which requirements they would need to meet. The model explains those results in plain language, highlights what is missing, and helps structure the personal report. In EduRelia, authorization decides that a parent may see their child's progress. The model helps phrase that progress in a supportive, age-appropriate way. In each case the model adds real value — clarity, structure, tone, speed — without being trusted with the facts. How we enforce the rule A rule that depends on everyone remembering it will eventually be broken. So we build it into how AI features are engineered. 1. Grounding: facts go in, not out When a feature needs a fact, the application retrieves it from the verified system first and passes it to the model as input. The prompt makes clear that the model must use only the provided data for that fact and must say so when data is missing. This inverts the usual demo pattern. Instead of asking "How many units are left?", the system asks the database, then asks the model: "Here is the current stock for these items. Write a short briefing for the store owner." 2. Structured outputs and schema validation For business-critical responses, we ask models to return structured data — for example, a JSON object with specific fields — rather than free text. That output is then validated against a schema before it is used. If a field is missing, has the wrong type or falls outside an allowed range, the response is rejected. Depending on the feature, the system retries, falls back to a simpler template, or routes the case to a person. We cover this in detail in Structured outputs: validating AI responses before users see them. 3. Cross-checking against the source Where it is cheap to do so, we compare what the model says with what the verified system holds. If a briefing mentions a product, does that product exist in the store? If a report lists a requirement, is it in the rules the engine returned? Mismatches are treated as failures. 4. Versioned prompts with owners Prompts are not strings buried in code that anyone can edit. Each important prompt has an identifier, a version, an owner and a change log. When it changes, we know who changed it, when and why. See Prompt versioning and evaluation for production AI. 5. Evaluation sets Before an important prompt or model change goes live, it is tested against a fixed set of examples with known correct behaviour — including tricky cases where the right answer is "I don't have that information". If quality drops, the change does not ship. 6. Fallbacks and degraded modes Models fail: they time out, return malformed output or become temporarily unavailable. Every AI feature has an approved degraded mode. A briefing can fall back to a plain list generated from the data. A report can be delivered without the narrative explanation. The facts are never lost when the model is. See Fallbacks and degraded modes for AI features. 7. Human escalation Some cases should not be handled automatically at all — unusual eligibility situations, sensitive messages, outputs that trip safety checks. Those go to a person. Designing that path is part of the feature, not an afterthought. See Designing human escalation paths in AI products. Why not just use a better model? Models improve quickly, and each generation hallucinates less than the last. It is tempting to wait for a model that "just gets it right". We do not think that is the right bet for business-critical facts, for three reasons. The model does not have your data. No matter how capable it is, a general model does not know your store's stock level this morning or the version of a visa rule that took effect last month. That information lives in your systems. Rules change. Visa requirements, school policies and store inventories change constantly. A verified, versioned database can be updated, audited and rolled back. A model's training cannot. Accountability matters. When a customer asks "why did you tell me this?", we want to point to a record: this rule, this version, this data. "The model said so" is not an answer we are willing to give. Better models make the explanation layer better. They do not remove the need for a source of truth. The same principle, applied to consumer products It might seem that entertainment products are exempt from this discipline. They are not — the facts are just different. In ZodiVela, readings are positioned as reflection and entertainment, not predictions or professional advice. That positioning is itself a rule the product must respect. The model is guided to write in that register, and outputs are checked so they do not drift into medical, legal or financial claims. What a user has paid for, which readings they have unlocked and what their account contains all come from the billing and account systems, never from the model. We discuss this further in Positioning consumer AI honestly. The architecture in one picture Most of our AI features share a simple shape: Request — a user action or scheduled job needs an AI-assisted output. Retrieve — the application fetches the relevant verified facts from the product's own systems. Prompt — a versioned prompt combines those facts with instructions about tone, format and limits. Gateway — the call goes through the AI gateway, which applies the product's model routing, budget, rate limits and logging. Validate — the output is checked against a schema and, where possible, against the source data. Deliver or fall back — valid output is shown; invalid output triggers a retry, a degraded template or a human. Observe — latency, cost, failures and quality signals are recorded per product. Nothing in that flow asks the model to be the source of truth. Common mistakes we avoid Over the years we have seen — and occasionally made — a few recurring mistakes with AI features. Letting the model fill gaps. If data is missing, the model should say so, not guess. Prompts and validation must enforce this. Trusting free text for decisions. If code needs to act on an AI output, that output must be structured and validated. Silent prompt edits. A one-word change in a prompt can change behaviour. Versioning and evaluation catch this. Shared prompts across products. A prompt tuned for one audience behaves badly with another. Each product owns its prompts. No budget limits. An AI feature without a spending cap is an invoice waiting to surprise you. See Cost discipline for AI products. No plan for failure. Every AI call can fail. The product must still work when it does. What this means for people using our products If you use an Oryvelon product with AI features, here is what the rule means for you: The numbers you see — stock, prices, eligibility results, account balances — come from the product's own systems. The explanations, summaries and suggestions are generated by AI and are designed to help you understand those numbers. When information is missing, the product should tell you rather than invent it. In sensitive or unusual situations, a person may be involved. We think that is what people reasonably expect, even if they never put it into words. How the rule shapes product design The rule does not only affect engineering. It changes how products are designed from the first sketch. When we scope a new AI feature, the first question is not "what can the model do?" It is "which facts does this feature depend on, and where do they live?" If the answer is "nowhere yet", building the source of truth becomes part of the project. For KeşifAtlası, that meant a rules database with versions and sources before any explanation layer. For MerchNivo, it meant a reliable connection to store data before any briefing. We describe the KeşifAtlası side in Rules engines for eligibility decisions. The second question is "what should the product do when the model is wrong or unavailable?" Answering that early produces better products: the degraded mode is often a clear, simple view of the underlying data, which turns out to be useful in its own right. The third question is "how will the user know which parts are facts and which are explanation?" Good interfaces make that visible — for example by showing the underlying numbers next to the narrative, or by labelling generated text as a summary. See Earning user trust in AI products. The review checklist we use for every AI feature Before an AI feature ships in any Oryvelon company, it goes through a short review. The questions are the same whether the feature is a store briefing, a visa report or a reading. Facts inventory. Which facts does the output contain, and which system is the source for each? Any fact without a source is either removed or given one. Missing-data behaviour. What does the feature say when a fact is unavailable? We test this explicitly with incomplete inputs. Output contract. Is the output structured where code depends on it? Is there a schema, and is it enforced? Worst plausible error. If the model gets this wrong in the worst realistic way, what happens to the user? The answer decides how much validation and human review the feature needs. Degraded mode. What does the user see if the model is slow, down or rejected by validation? Prompt ownership. Who owns the prompt, where is it versioned and which evaluation set covers it? Cost ceiling. What is the budget for this feature per day and per month, and what happens when it is reached? Data boundary. Does any data from another product, or any data the user did not provide for this purpose, reach the model? The answer must be no. Labelling. Can the user tell which parts are data and which are generated explanation? Monitoring. Which signals will tell us the feature is degrading — failure rate, validation rejections, complaints, latency? A feature that cannot answer these questions is not ready, however impressive the demo looks. In our experience the checklist rarely slows a team down for long; it mostly moves hard conversations earlier, when they are cheap. What we learned by applying the rule A few lessons have come up repeatedly as we applied this rule across very different products. Users trust the product more, not less, when AI admits gaps. A briefing that says "I couldn't read last night's orders, here is what I can see" earns more confidence than one that glosses over the problem. The data layer often turns out to be the real product. Building a clean, verified source of truth — a store model, a rules database, a consent record — creates value on its own. The AI layer makes it accessible; the data layer makes it correct. Smaller models go further than expected. When facts are supplied rather than recalled, many tasks become formatting and explanation problems, which cheaper, faster models handle well. That keeps costs predictable. See Model routing. A worked example: a refund spike in MerchNivo The rule is easiest to see in a situation where both layers matter. Imagine a store where refunds on one product rise sharply over three days. What the verified data decides. The store's order and refund records, read through the Shopify API, establish the facts: which product, how many refunds, over what period, compared with the store's own recent baseline. The threshold that turns this into an alert is a rule written in code, not a judgement made by the model. Stock and inventory figures come from the store as well; we explain why that matters in Inventory truth. What the AI explains. The model receives those facts and writes the part of the morning briefing a busy owner will actually read: refunds on this product are up, most refund notes mention sizing, here are two things worth checking — the size chart on the product page and whether a recent variant change caused confusion. It can draft a clearer size note for the owner to review. What the AI is not allowed to do. It does not state a refund total that the data did not contain. It does not claim a cause as fact; it says "most notes mention sizing" because that is what the notes say. It does not issue refunds, change prices or pause the product. Those are suggested actions a person approves. If the model call fails, the owner still gets the signal as a plain line: product, refund count, period. The explanation is lost for a morning; the fact is not. We describe how signals like this are defined in Shopify operations signals. The same split applies when a visa rule changes in KeşifAtlası: the rules database is updated with its source and effective date, and the explanation layer picks up the new result without anyone editing a prompt to "know" the new rule. Summary AI is one of the most useful tools we have for building products with small teams. It is also one of the easiest to misuse. At Oryvelon we use it for what it does best — explaining, summarising, drafting, structuring — and we keep business-critical facts in systems built to hold them. That separation is enforced through grounding, structured outputs, validation, versioned prompts, evaluation, fallbacks and human escalation. It is the same across every company in the group, and it is one of the main reasons those companies can be trusted. To see how each company applies it, visit the Companies page. For the broader operating model, see Build. ## Shared infrastructure, separate data URL: https://oryvelon.com/insights/shared-infrastructure-separate-data/ Category: Data, Privacy & Security Whenever one organisation runs several products, an uncomfortable question follows close behind: does it combine what it knows about people? For many groups the honest answer is "yes, eventually", because combined data is tempting and the tools to combine it are easy to use. At Oryvelon the default answer is no. We share infrastructure across our companies. We do not share the people who use them. This note explains what that means in practice: where the boundaries sit, how they differ between companies, why consent is handled the way it is, and what each product has to plan for before it goes live. Two very different kinds of sharing It helps to separate two ideas that often get mixed up. Sharing infrastructure means using the same standards, tools and practices to run several products: the same approach to deployment, monitoring, security headers, domain management, email authentication, AI model access and cost tracking. Nothing about a customer moves between products. What moves is know-how. Sharing data means that information about a person collected by one product becomes available to another: a purchase history informing a different product's recommendations, an email address added to another brand's newsletter, a profile enriched with behaviour from elsewhere. The first makes every company stronger. The second quietly changes the promise each company made to its customers. We do a lot of the first and none of the second by default. Our default: no cross-product user graph A "user graph" is the web of connections an organisation builds between accounts, devices, emails and behaviour. Large platforms build enormous ones. For a group like ours, a cross-product graph would mean linking a person's activity in MerchNivo, ZodiVela, KeşifAtlası and so on into a single profile. We do not build one. Each company keeps its own users in its own database. If the same person signs up for two products, they have two accounts, and nothing in our systems joins them. Could there be a future case where two specific products benefit from a connection — for example, a merchant who explicitly wants to link their MerchNivo workspace with a store run by our services company? Possibly. If so, it would be designed as an explicit, consented link between two independent systems, visible to the user and reversible. It would not be a background join that nobody asked for. Boundaries by company Not every product needs the same boundary. The sensitivity of the data decides how strict the rules are. Here is how it looks today. Company Data kept separate Special rule MerchNivo Shopify stores, orders, products, store customers Store-by-store tenant isolation CastLyra Talent profiles, business briefs, messages Explicit consent and business access rules EduRelia Student, parent, teacher and school data School tenants, row-level security, child-data minimisation ZodiVela User profiles, readings, uploads No cross-product profiling KeşifAtlası Assessments, income bands, family and application details Minimised sensitive data, versioned rules Noveniq Commerce and order data Independent production business Sinem Keser Beauty Academy Appointments, students, purchases Independent brand and customer relationship A few of these deserve more explanation. MerchNivo: store-by-store isolation MerchNivo reads data from Shopify stores. That data belongs to each merchant, and it includes information about the merchant's own customers. Every store is a separate tenant: one merchant can never see another's data, and store data is never used for any other Oryvelon product. Even Noveniq, which the group owns and which serves as a real testing environment for MerchNivo, is treated as just another tenant with its own boundary. See Using a real store as a product test-bed. EduRelia: the strictest standard in the group EduRelia handles information about children. That puts it at the top of our sensitivity scale. Schools are separate tenants, access is enforced with row-level security in the database rather than only in application code, and the product collects the minimum information needed to support learning. Student data is never used in any consumer product or for marketing. We describe the approach in Child-data minimisation in education technology. CastLyra: consent as a product feature On a talent marketplace, visibility is the product — but it has to be controlled by the talent. CastLyra's rules decide which businesses can see which details, and consent is explicit and recorded. Talent data is never merged with any other Oryvelon product. See Verification as the foundation of a talent marketplace. KeşifAtlası: minimise what is sensitive People researching visas and relocation may share income bands, family circumstances and plans. KeşifAtlası collects only what the eligibility rules genuinely need, and assessment data does not flow automatically into marketing profiles of any product. Marketing consent is not product consent This is the part of data handling where good intentions most often slip. When someone signs up to use a product, they agree to that product processing their data to provide the service. That is product consent (or, depending on the legal basis, a contractual necessity). It is specific to the product. When someone agrees to receive marketing, that is marketing consent. It is a separate decision, and it is also specific: agreeing to hear from ZodiVela about new reading formats is not agreeing to hear from KeşifAtlası about visa reports. At Oryvelon the two are recorded and handled separately, per company. Using one company's customer list to promote another company is off the table unless the person has explicitly agreed to it. We cover the practicalities in Marketing consent is not product consent. Shared AI without shared data AI raises the stakes, because it is easy to accidentally send data to places it should not go. Every AI feature in the group uses the same kind of gateway for model routing, budgets, logging and fallbacks. But each product has: its own AI project and API keys; its own prompts, owned and versioned by that product; its own knowledge base for retrieval, if it uses one; its own budget and cost reporting; its own logs, filtered for personal information. The gateway standardises how we call models. It never merges what products send to them. A store summary from MerchNivo and a reading from ZodiVela never share a prompt, a knowledge base or a log. See Retrieval without leakage: private knowledge bases per product and What an AI gateway does. Analytics without a single identity We still need to understand how people move between our companies — for example, how many visitors to oryvelon.com go on to visit a company site. We do that with consistent campaign parameters on links between sites, and by looking at aggregated, product-level numbers. We do not stitch individual users together across products. See Analytics without surveillance and UTM standards for measuring traffic between sister brands. What every product plans for from day one Boundaries are easiest to build at the start. Before any Oryvelon product goes live, its data plan answers five questions. Minimum collection. What is the smallest set of data that lets the product work? Everything else is left out until there is a clear reason. Role-based access. Which roles exist — customer, admin, teacher, parent, business, talent — and what can each one see? Access is defined by roles, not by individual exceptions. Auditability. Which actions are critical enough to log — changes to permissions, exports, deletions, sensitive views — and where is that log kept? Retention. How long is each kind of data kept, and what triggers its deletion? Export and deletion. How does a person get a copy of their data, and how do they have it deleted? We explain the last two in more depth in Retention, export and deletion. What this costs us — and why it is worth it Separate data is not free. It means: separate databases to operate, back up and monitor; separate accounts for people who use more than one product; separate analytics setups; more careful engineering whenever a feature touches personal data; saying no to some marketing ideas that would be easy with combined data. We accept those costs because the alternative is worse. Combined data creates a single point of failure: one breach, one misconfigured query or one overreaching campaign affects every company at once. It creates regulatory complexity, because each product's legal basis for processing is different. It makes it much harder to ever separate a company from the group. And, most importantly, it changes the relationship each company has with its customers in a way they did not agree to. How the boundaries are checked over time Boundaries that are drawn once and never checked tend to erode. A convenient export here, an extra field there, a new vendor that receives more than it needs. So the boundaries are part of our regular operating rhythm rather than a one-off design exercise. At design time, every new feature that touches personal data answers the five planning questions again: what it collects, who can see it, what is logged, how long it is kept and how it is removed. At vendor selection, any new tool that will receive personal data is reviewed for what it actually needs. Analytics, email, support and AI providers are configured per company, with the minimum data required. At access reviews, which happen every quarter, we check who can reach each company's data and remove access that is no longer needed. See Least-privilege access for small teams and contractors. At incident review, if something goes wrong — a misrouted email, an over-broad export, a log that captured more than it should — the fix includes a change to the standard so the same mistake is harder to make in every company. This rhythm is not glamorous, but it is what turns a principle into a habit. Requests we decline, and what we do instead Principles become clearer when you see the requests they rule out. Here are the kinds of ideas that come up in any multi-brand group, and how we handle them. "Let's email all our customers about the new product." We do not send one company's customers marketing for another company. Instead, the new product is announced on oryvelon.com, on the founders' own channels and — where it genuinely fits — in the existing product's own communication about its own features. Anyone who has explicitly opted in to hear about the group can be told directly. "Let's enrich profiles with data from our other apps so recommendations are better." Recommendations in each product use that product's data. If a product needs more information to be useful, it asks the user for it, explains why and lets them decline. "Let's use one analytics property for everything so we can see the whole funnel." Each company has its own measurement. To understand movement between companies, we use campaign parameters on links and compare aggregated numbers. We do not track individuals across brands. "Let's train one model on everything we know." Each product's knowledge base, prompts and evaluation data stay with that product. Techniques that work well are shared as documentation; data is not. "Let's give the support team access to all products so they can help faster." Support access is granted per company and per role. A person helping CastLyra talent does not automatically see EduRelia records. Access is reviewed regularly and removed when it is no longer needed. Each of these declines has a small cost in convenience. Together they are what makes the group trustworthy. Five questions before combining any data Occasionally there may be a genuine case for connecting data between two companies. When that happens, we ask five questions before anything is built. Did the person clearly ask for it? A connection the user requested and can see is very different from one done in the background. Is the purpose specific and explained? "Better experience" is not a purpose. "Show your store's sales in your agency report" is. Is it the minimum data needed? Connecting two systems does not mean copying everything from one to the other. Can the person undo it? Every connection should be reversible, with data removed when it is. Would we be comfortable describing it publicly? If we would hesitate to explain it on this website, we should not build it. If any answer is uncomfortable, the connection is not built. A note for partners and customers If you are a business working with one of our companies, the practical implications are simple. Your data is held by that company, under that company's privacy notice, and used to provide the service you signed up for. It is not shared with other Oryvelon companies, and it is not used to market their products to you or your customers. If you are an individual using one of our products, the same applies. You can ask that company for a copy of your data or for its deletion, and the request concerns that company alone. A worked example: one person, three companies Consider a hypothetical person — call her a small business owner — who uses three Oryvelon companies over a year. She runs a Shopify store and connects it to MerchNivo. Later she books a course at Sinem Keser Beauty Academy. Later still she uses KeşifAtlası to check whether she could move abroad. Inside the group, those are three customers. MerchNivo holds her store connection inside her store's tenant. The academy holds her booking and payment record. KeşifAtlası holds her assessment answers, including income band and family details, collected only because the eligibility rules need them. Nothing links the three records, and nobody in the group is supposed to know they belong to the same person. Here is what that means in practice: The KeşifAtlası assessment never informs what MerchNivo suggests, and her store's sales never affect how KeşifAtlası treats her. Unsubscribing from the academy's newsletter changes nothing in the other two, because each consent was given separately. If she asks KeşifAtlası to delete her data, that request is handled by KeşifAtlası. Her MerchNivo workspace is untouched, because it belongs to a different relationship. If one of these companies were ever sold, her records would go with that company only. That last point is the test we find most useful. If a company cannot be separated without first untangling its customers from everyone else's, the boundary was never real. Common ways data leaks between products by accident Few groups decide to merge customer data. Most drift into it through tooling. These are the leaks we design against. One tracking pixel or tag container across brands. A shared advertising or analytics tag quietly builds the cross-brand profile a group said it would not build. Each company has its own tags and properties. One email tool account with one audience list. If every brand's subscribers sit in one list with a "brand" field, one wrong filter sends the wrong people the wrong campaign. Audiences live in separate workspaces. Production data copied into test environments. A realistic test database is tempting, and it is how real customer records end up on a laptop. Test data is synthetic or anonymised. See Local, preview, staging, production. Logs that carry whole payloads. A debug log that stores full AI requests can hold personal information from one product in a place other people can read. Logs are filtered and stay per product. Shared database users. Even with separate tables, a single database user with access to everything removes the boundary where it matters most. Isolation is enforced at the data layer, as described in Tenant isolation explained. None of these requires bad intent. That is why the standards cover them explicitly. Summary Oryvelon exists to share foundations across companies. Data is deliberately not one of those foundations. Each company holds its own users, its own records and its own promises, and the group's job is to make sure those promises are kept — with tenant isolation where it is needed, the strictest standards where children are involved, explicit consent where visibility is the product, and separate handling of marketing permissions everywhere. Share the plumbing, not the people. It is one of the simplest rules we have and one of the most important. For how each company applies it, see the Companies page. For the wider operating model, see Build. ## How we start a new company: our 10-step protocol URL: https://oryvelon.com/insights/how-we-start-a-new-company/ Category: Company Building Most ideas should not become companies. That is not pessimism; it is arithmetic. For every idea worth building there are many that sound good at dinner and fall apart when you ask who pays, how often and why they would keep paying. A company builder has a particular risk here. Because we can launch quickly, we could also launch too much. A portfolio of half-built brands, each with a nice logo and no customers, is worse than having no portfolio at all. So every new Oryvelon company goes through the same protocol, in the same order. It is ten steps long. Most ideas stop somewhere in the first four — which is exactly what the protocol is for. Why a protocol at all? A protocol does three things that good intentions do not. It forces the hard questions early. The expensive mistakes in a new company are usually decided in the first weeks: who the customer is, how money flows, how sensitive the data is. A protocol puts those questions before any code or design. It makes ideas comparable. When every idea is written up the same way, it becomes possible to compare them honestly and decide where attention should go. It makes the group's knowledge reusable. Every company that goes through the protocol leaves behind a clean record of how it was set up. The next company benefits. The ten steps 1. Problem and target user, in one sentence We start by writing the problem and the person who has it in a single sentence. Not a paragraph, not a pitch deck — one sentence. For MerchNivo it was roughly: independent Shopify merchants run their stores across dozens of screens and nobody watches the whole picture. For KeşifAtlası: people planning to study, work or live abroad face scattered information, and generic AI answers about visas can be confidently wrong. If we cannot write the sentence, we do not understand the problem well enough to build anything. If the sentence describes a trend rather than a person with a problem, we stop. 2. Revenue mechanism Next: who pays, for what, and why? This is written before the brand, the design or the product. The answers vary across the group. MerchNivo is a subscription with usage-based components. CastLyra combines business subscriptions, premium plans for talent and, over time, booking commission. EduRelia has family subscriptions and annual school licences. ZodiVela uses subscriptions, credits and one-off premium readings. KeşifAtlası sells paid personal reports and works with partners. Noveniq earns product margin. What they have in common is that the answer is concrete. "We'll figure out monetisation later" is not an answer that passes step two. 3. Recurring income loop A company that has to re-win every customer every month is exhausting to run. So step three asks: how does revenue repeat without constant manual effort? For a subscription product the loop is obvious, but it still has to be designed — why would someone stay subscribed in month six? For a report-based product like KeşifAtlası, the loop might come from new situations (a family's circumstances change, rules change), partner referrals or adjacent services. For a commerce brand, it comes from repeat purchases and a reason to return. We cover this step in more depth in Designing a recurring revenue loop from day one. 4. Brand and domain ownership Only now do we think about the name. The brand must be available, the domain must be owned by the group, and there must be no obvious conflict with an existing company in the same category. Domains are registered with auto-renew and domain lock enabled, under accounts protected by two-factor authentication. Losing a domain through an expired card or a compromised account is one of the most avoidable disasters in digital business. See Domain, DNS and email security for a multi-brand group. 5. A source-of-truth document Every company gets a written source of truth: what it is, who it serves, how it earns, what it will not do, how it relates to the group and which data rules apply. Later documents — product specs, marketing briefs, technical plans — must agree with it. This sounds bureaucratic. In practice it saves enormous time. When someone joins a project, asks "why is it like this?" or proposes a change that contradicts the original intent, the answer is written down. See Why every company starts with a source-of-truth document. 6. Naming to standard Repositories, deployment projects, AI projects and analytics properties are named according to a group standard. It is a small step that pays off every day: anyone can find anything, and nothing gets mixed up between companies. Separate AI projects per company also mean separate keys and budgets from the start. 7. Data sensitivity classification Before a line of product code is written, we classify the data the company will hold and decide how strictly it must be protected. A consumer entertainment app, a marketplace with private contact details and a platform holding children's learning records sit at very different points on that scale. The classification decides tenant isolation, access rules, logging, retention and whether additional review is needed for AI features. Doing this early is the single biggest cost saving in the protocol. Retrofitting privacy into a product that was not designed for it is slow, risky and expensive. See Tenant isolation explained. 8. A phase-zero site The company gets a real website before the product is finished: real content explaining the problem and the solution, a way to join a waitlist or register interest, proper structure for search engines and AI assistants, and measurement. The phase-zero site tests whether the message resonates, starts building an audience and gives the brand a home. It is not a placeholder page. See Phase-zero sites: launching a brand before the product exists. 9. MVP acceptance criteria We define what "done" means for the first release. Not a feature wish-list — a short list of conditions that must be true before real customers use it. For an AI product, those conditions include the facts inventory, degraded modes and budget limits described in AI explains, verified data decides. Clear criteria stop the first version from growing endlessly, and they make launch a decision rather than a feeling. 10. 30/60/90-day KPIs and a decision Finally, we write down what success looks like at 30, 60 and 90 days after launch, and we commit to making an explicit decision at those points: continue, stop or scale. The KPIs differ per company — sign-ups, activation, paid conversion, retention, marketplace liquidity, report sales, repeat purchases — but the commitment to decide is the same. See Continue, stop or scale. Why the order matters The protocol is not just a checklist; the sequence is deliberate. Revenue before branding. It is very easy to fall in love with a name and a logo. Putting revenue first means we only spend time on brands for ideas that have a plausible business behind them. Data classification before code. The architecture of a product that holds children's data is different from one that holds horoscope preferences. Deciding late is expensive. A public site before a finished product. Real feedback from real visitors arrives early, while the product can still change direction cheaply. Decision points before launch. Writing KPIs before launch stops us from redefining success afterwards to match whatever happened. What does not appear on this website Ideas stay internal. A company is listed on the Companies page only when it has passed the protocol and exists as a real brand with its own website. This keeps oryvelon.com honest: what you see is what we operate. Where most ideas stop It is worth being explicit about where ideas usually fall out of the protocol, because the pattern is instructive. Step 1 stops ideas that are really trends. "An AI app for X" is not a problem statement. When we try to write the sentence and find ourselves describing a technology instead of a person, the idea goes back on the shelf. Step 2 stops ideas without a payer. Many useful things are hard to charge for. If the only plausible revenue is advertising to an audience we do not yet have, or a vague "enterprise version later", we stop. Step 3 stops one-off businesses. Some ideas have a clear payer but no reason for them to return. Those may still be good businesses for someone, but they rarely fit a group that wants companies to compound. Step 7 reshapes ideas rather than stopping them. Data classification occasionally reveals that an idea would require holding information we are not prepared to hold responsibly. More often it changes the design: collecting less, isolating more, or removing a feature that needed sensitive data it did not really need. By the time an idea reaches step eight, it has usually changed shape at least once. That is a good sign. It means the protocol is doing its job. A worked example, step by step To make the protocol concrete, here is a condensed version of how an idea moves through it. The details are simplified, but the sequence is the one we follow. Imagine a problem we hear repeatedly from people around us: creative professionals — models, photographers, stylists — struggle to be found by serious businesses, and businesses struggle to know whether a profile is genuine. Step 1 turns that into one sentence with a person in it: brands need verified creative talent, and talent needs trustworthy requests. Step 2 asks who pays. Businesses searching for talent can pay a subscription; talent who want more visibility can pay for premium features; later, completed bookings can carry a commission. Step 3 asks why revenue repeats. Brands produce content continuously, so their need for talent recurs. Talent keep profiles current as their careers develop. Step 4 secures a name and domain — in this case, CastLyra. Step 5 writes the source of truth: a global marketplace, verification at its core, talent in control of visibility, no merging of talent data with any other product. Step 6 names the repositories, AI project and analytics property according to the group standard. Step 7 classifies the data: personal profiles, private contact details, messages. That leads directly to explicit consent rules and business access controls. Step 8 publishes a phase-zero site that explains the idea to both sides of the marketplace and invites them to register interest. Step 9 defines MVP criteria: talent can apply and be verified, businesses can search and send a request, contact details are only shared under the access rules, and every AI-assisted match can be explained. Step 10 sets the numbers that matter for a young marketplace — verified profiles, active businesses, requests sent and answered — and fixes the dates on which we will decide. The result is a company that knows what it is before it writes much code. We explore the marketplace-specific decisions in Verification as the foundation of a talent marketplace. Common objections to working this way We sometimes hear that a protocol like this slows things down. Our experience is the opposite, but the objections are worth answering. "Great companies come from intuition, not checklists." Intuition is how ideas start. The protocol does not replace it; it tests it. A strong intuition survives ten questions easily. "Writing a source of truth is wasted time when everything will change." Things will change, and the document changes with them. What matters is that there is one place where the current intent is written down, so changes are deliberate. "A phase-zero site before the product is premature." It would be, if it were a placeholder. A real phase-zero site with useful content often becomes the company's best long-term marketing asset, because it starts earning search visibility months before launch. "Committing to a stop decision is demoralising." It is more demoralising to spend a year on something that should have been stopped in month three. Clear decision points protect the team's time and the group's capital. What the protocol leaves behind When a company finishes the protocol, it has a set of assets that make everything afterwards easier: a one-sentence problem statement everyone can repeat; a documented revenue mechanism and recurring loop; an owned, secured brand and domain; a source-of-truth document; consistently named repositories, AI projects and analytics; a data classification with matching technical controls; a live phase-zero site with real content and measurement; written MVP acceptance criteria; dated 30/60/90-day KPIs and a decision commitment. Those assets are also what make a company separable later. Everything a partner, investor or acquirer would ask for exists from the beginning. See Designing every company so it could stand alone. How the protocol itself evolves The protocol is not fixed. Each company that goes through it teaches us something, and the lessons are folded back in. Some changes are small — a better template for the source of truth, a clearer naming convention. Some are more significant. The emphasis on data classification before code, for example, became sharper after seeing how differently an education product and a consumer product needed to be built. The requirement for AI degraded modes in MVP criteria came from seeing how often AI calls fail in ordinary operation. We review the protocol alongside our operating cadence, and changes apply to the next company that starts. How the steps weigh differently by company type Every company goes through all ten steps, but the steps do not carry equal weight for every company. Knowing where the real work sits for a given business saves time on the steps that are routine and focuses attention on the ones that decide the outcome. Company type Where the hard work usually sits Software for businesses (MerchNivo) Steps 2 and 9: pricing that tracks usage, and MVP criteria that guarantee every number comes from the store Marketplace (CastLyra) Steps 3 and 7: a reason for both sides to return, and consent rules for contact details Education (EduRelia) Step 7 above all: data about children shapes the whole architecture, and the buyer is often a school rather than the user Decision support (KeşifAtlası) Steps 5 and 9: what the product will never do, and how hard questions reach a person Commerce (Noveniq) Steps 2 and 3: margin after every cost, and a reason customers buy again For EduRelia, for example, step seven is not a checkbox. It decides what the product may ask a student, what a school administrator can see, how long records are kept and which AI features need extra review. We describe that approach in Child data minimisation in edtech. For Noveniq, the same step is comparatively light, while the revenue questions are unforgiving: a store that sells well at a loss is not a business. The lesson for anyone using the protocol is to spend a few minutes before starting to mark which two or three steps are the real risk for this particular idea. Those steps get the most time and the most scepticism. Mistakes we watch for inside the protocol Having a protocol does not guarantee it is used well. These are the ways it goes wrong, including ones we have had to correct in ourselves. Filling in steps after the fact. The protocol only works if each step is answered before the next begins. Writing the revenue mechanism once the product is half-built turns a decision into a description. Copying a data classification from another company. It is tempting to reuse the classification from a similar product. Similar products can hold very different data. Each company gets its own classification, even if it ends up close to another one. KPIs nobody can measure. Step ten sometimes names numbers that the phase-zero site and the product were never set up to record. We now check at step eight that every KPI in step ten has a matching event or data source. Treating naming as a creative exercise first. A beautiful name with an unavailable domain or an obvious conflict wastes days. Availability and ownership are checked before anyone gets attached. Skipping the stop conversation. When an idea reaches step ten, the team is invested. That is precisely when the stop criteria need to be written most plainly. How we weigh those decisions across the group is covered in Capital allocation inside a company builder. Using the protocol yourself You do not need to be a company builder to use these steps. If you are a founder with one idea, working through them honestly — especially steps one, two, three and seven — will save you time and money whether or not you end up building. A simple way to start is to open a blank document with ten headings and try to fill each one in a paragraph or less. Where you cannot, you have found the work that matters most. Summary Every Oryvelon company starts the same way: a one-sentence problem, a concrete revenue mechanism and recurring loop, an owned and secured brand, a written source of truth, standard naming, data classification before code, a real phase-zero site, written MVP criteria and dated 30/60/90-day decisions. The order is deliberate, the hard steps differ by company type, and most ideas stop early — which is the protocol working as intended, not failing. ## Cost discipline for AI products URL: https://oryvelon.com/insights/cost-discipline-for-ai-products/ Category: AI Operations There is a pattern we have seen many times in AI products. A feature ships. It works well. Usage grows. Nobody looks closely at the bill until a month when it is suddenly the largest line item — larger than hosting, larger than the database, sometimes larger than the revenue the feature brings in. AI cost is unusual because it scales with use in ways that are easy to underestimate. Every request costs money. Longer inputs cost more. Longer outputs cost more. Retries cost more. A prompt change that adds a paragraph of context can quietly raise costs across every call. In a group of companies, there is an extra risk: one product's spending can hide inside another's, or inside a combined invoice that nobody breaks down. At Oryvelon we treat AI cost as a first-class metric for every company, with the same seriousness as revenue. Principle: every product sees its own economics The foundation of cost discipline is visibility. For each company we aim to track, side by side: hosting and infrastructure; database and storage; AI model usage; email and messaging; third-party tools and subscriptions; revenue and gross margin. A single group invoice must never hide whether an individual product is healthy. If ZodiVela's AI usage doubles, we want to see it in ZodiVela's numbers, not as a slightly larger total that nobody can explain. See Unit economics per product in a shared-infrastructure group. Separate AI projects, keys and budgets The simplest structural decision we made is also one of the most effective: each product has its own AI project, with its own API keys and its own budget. That does three things: Spend is attributable. Every request is billed to the product that made it. Limits are local. A soft limit warns the team when spending approaches the monthly budget; a hard limit stops runaway usage before it becomes a financial problem. One product hitting its limit does not affect the others. Keys are contained. If a key leaks, the damage is limited to one product and the key can be rotated without disrupting the group. The shared AI gateway applies these budgets consistently, but the budgets themselves belong to each company. Model routing: the right model for the task Not every task needs the most capable model. Classifying a support message, extracting fields from a product description, summarising a short text or reformatting data can often be done well by smaller, faster, cheaper models. Complex reasoning — interpreting an unusual combination of eligibility rules, or writing a nuanced explanation — may justify a stronger one. We route tasks deliberately: Task type Typical routing Classification, tagging, extraction Small, fast model Short summaries and reformatting Small or mid-sized model Explanations grounded in verified data Mid-sized model, stronger if quality requires Complex multi-step reasoning Strong model, used sparingly Because our products ground models in verified data rather than asking them to recall facts, many tasks become formatting and explanation problems — exactly the kind smaller models handle well. We explain the routing logic in Model routing: matching tasks to the right AI model. Caching where it is safe If the same input produces the same useful output, there is no reason to pay for it twice. We cache responses for deterministic, low-variance tasks: a description of a product that has not changed, an explanation of a rule version that has not changed, a classification of identical text. Caching needs care. It must never serve one user's personalised output to another, and it must be invalidated when the underlying data changes. For that reason, cache keys include the product, the prompt version and a fingerprint of the input data. See Latency, caching and batching in AI features. Batching work that is not urgent Some AI work does not need an instant answer: nightly store analyses, bulk content suggestions, periodic report generation. Running that work in batches — often through background jobs — lets us use cheaper processing options where providers offer them, smooth out load and retry failures without a user waiting. Output limits Outputs cost money too. A briefing that could be five bullet points does not need to be five paragraphs. We set output length limits per feature and design prompts to ask for the format we actually need. Shorter outputs are cheaper, faster and usually easier to read. Watching for silent cost growth The most dangerous cost increases are the ones nobody decides to make. A few examples we watch for: Prompt creep. Each change adds a little context or a few instructions. After a dozen changes, every request is much larger. Versioned prompts make this visible. Retry storms. A validation rule that rejects many outputs can cause repeated retries. Monitoring rejection rates catches this. Context bloat. Retrieval systems that send too much text to the model on every request. We cap retrieved context and measure its size. Forgotten features. Background jobs that keep running long after anyone uses their output. Tools, reviewed quarterly AI is not the only cost that creeps. Shared infrastructure only saves money if it stays lean, and SaaS subscriptions have a way of multiplying: a monitoring tool here, a form builder there, two overlapping analytics services because two people signed up for trials. Every quarter we review the tools and subscriptions used across the group. For each one we ask: who uses it, which companies depend on it, whether another tool already does the same job, and whether its cost is attributed to the right company. Duplicates are closed. Unused seats are removed. See The quarterly tool review: fighting SaaS sprawl. A cost review for a single AI feature When we look at the cost of an individual AI feature, we walk through the same short list. Volume. How many requests per day, and is that growing in line with usage or faster? Size. What is the average input and output size, and has it changed since the last review? Model mix. Which models handle which share of requests, and could any share move to a cheaper model without hurting quality? Cache hit rate. For cacheable tasks, how often are we serving a stored answer, and could more be cached safely? Failure cost. How much spending goes on requests that fail validation or time out and are retried? Value. What does the feature contribute — conversions, retention, time saved for users — relative to what it costs? The last question matters most. A feature that costs a lot and drives a lot of value is fine. A feature that costs a little and nobody uses is not. Quality and cost are not opposites It is tempting to see cost control as a trade-off against quality. In our experience, disciplined cost work often improves quality. Grounding makes smaller models viable. When a model receives verified facts and only has to explain them, a smaller model often writes as clearly as a larger one — and is less likely to wander. Shorter outputs are easier to read. A tighter output limit usually produces a more useful briefing. Batching makes retries calm. Background work can be retried and validated carefully without a user waiting, which reduces pressure to accept a borderline answer. Evaluation sets protect both. When we try a cheaper model or a shorter prompt, the same evaluation set tells us whether quality held. We only keep changes that pass. How it looks across the group Each Oryvelon company uses AI differently, so the cost profile differs too. MerchNivo runs regular analyses of store data and on-demand requests from merchants. Batching and caching matter most here, because store data changes on a known rhythm. KeşifAtlası generates explanations for eligibility results. Rules are versioned, so explanations for the same rule version and the same answers can be reused safely. ZodiVela produces personalised consumer readings. Output limits and model routing keep per-reading cost predictable as usage grows. EduRelia supports learning interactions, where latency and safety matter as much as cost. Budgets are set conservatively and reviewed often. CastLyra uses AI to assist matching between briefs and talent profiles. Most matching work can use smaller models over structured profile data. The tools are the same; the settings are tuned per company. Mistakes to avoid If you are building AI features yourself, these are the mistakes we see most often: one API key shared across several products, making spend impossible to attribute; no hard budget limit, so a bug or abuse becomes an expensive surprise; the strongest model used for every task by default; no measurement of input size, so prompt and context growth goes unnoticed; caching added carelessly, serving personalised output to the wrong user; AI cost reported only as a group total, never per product or per feature. Each of these is easy to fix early and painful to fix late. Setting a budget for a new AI feature New features are where cost surprises start, so every AI feature gets a budget before it launches, not after its first invoice. We estimate three numbers: expected requests per day, average tokens per request (input plus output) and the price of the model tier we plan to use. Multiplying them gives a daily cost; multiplying again gives a monthly figure. We then add a margin for retries and growth and set two thresholds. The soft limit sits around the expected monthly cost. Crossing it sends an alert to the team that owns the feature. Nothing breaks; someone looks. The hard limit sits well above the soft limit but below the point where the feature stops making economic sense. Crossing it switches the feature to its degraded mode until someone decides what to do. The estimate is always wrong in the details, but writing it down forces useful questions. Is the input really that large? Do we need that model? What happens at ten times the usage? It also gives the team a baseline to compare against once real traffic arrives. Pricing AI features for customers Cost discipline is not only about spending less. It is also about charging sensibly for what AI costs to provide. For products where AI usage varies a lot between customers, a flat subscription can become unfair in both directions: light users subsidise heavy users, or heavy users become unprofitable. That is why several Oryvelon companies combine a subscription with usage-based elements or credits. MerchNivo pairs a subscription with usage components; ZodiVela uses credits and one-off premium readings alongside subscriptions. The goal is not to maximise revenue per request. It is to make sure the price of a product moves roughly with the cost of serving it, so growth makes the business stronger instead of more fragile. We cover the pricing logic in Subscription plus usage: pricing logic for AI SaaS. Reporting AI cost in a way people read A cost report that nobody reads does not control anything. We keep AI cost reporting short and consistent, per company: total AI spend this month and last month; spend per feature, with the top three highlighted; cost per active user or per unit of value (per report, per store, per reading); any feature that crossed a soft limit, and why; one proposed change, if there is one worth making. Five lines, every month, for every company. It is enough to catch problems early and dull enough that it actually gets done. Choosing providers and models with cost in mind Model prices change often, and new models appear constantly. Rather than chasing every announcement, we evaluate providers and models against a stable set of questions. Does it pass our evaluation sets? A cheaper model that fails the tests for a feature is not cheaper; it moves the cost to our users and our support time. Is the price predictable? We prefer pricing we can model in advance. Features with unpredictable cost need tighter limits. Can we separate projects and budgets? A provider must let us give each company its own project, keys and limits. Without that, per-product visibility breaks. What happens when it is unavailable? Every provider has outages. The gateway needs an approved fallback — sometimes a second provider, sometimes a simpler non-AI path. What does it do with our data? Inputs may contain customer information. Data handling terms, retention and training policies matter as much as price. Is switching realistic? Because every call goes through the gateway with versioned prompts and evaluation sets, switching models is a configuration change followed by a test run — not a rewrite. That keeps us free to move when prices or quality change. This approach means we rarely make dramatic model changes. Instead, costs drift down steadily as individual features move to better-priced options that pass their tests. A worked example: the cost of a daily briefing Abstract advice about AI cost is easy to agree with and hard to apply. So here is how we think through a single feature, using a daily store briefing of the kind MerchNivo produces. The figures below are illustrative, not measurements. The naive version sends the model everything: yesterday's orders, the product catalogue, inventory levels, refunds and traffic, and asks it to "summarise what matters". For a store with a few thousand products, that input is large, and it is paid for every day, for every store. It is also worse for quality. The model is being asked to do arithmetic and spot trends in raw data, which is exactly where language models make confident mistakes. The version we prefer splits the work. Code computes the facts. Revenue compared with the same weekday last week, products close to selling out, a refund rate above the store's normal range, a sudden drop in conversion. These are ordinary calculations on the store's own data. Code decides what is worth mentioning. Thresholds and rules pick the handful of signals that crossed a line today. On a quiet day that list may be short. The model explains. It receives a compact, structured list of verified facts — perhaps a few hundred tokens instead of tens of thousands — and writes a short briefing with suggested actions. A schema checks the output. The response must match a defined structure, and every number in it must appear in the input. See Structured outputs and schema validation. Imagine a store whose refunds spike on one product line overnight. In the naive version, the model might notice, or might bury it in paragraph three. In the split version, the refund calculation flags it, the briefing leads with it, and the number shown is the store's own. The cost per briefing drops sharply, because input size is the main driver, and the briefing becomes more trustworthy at the same time. The general rule: do the arithmetic in code and pay the model for language. It applies well beyond commerce. When we spend more, not less Cost discipline is not a race to the cheapest option. There are places where we deliberately accept a higher cost per request, and being explicit about them keeps cost reviews from eroding things that matter. Safety checks stay on. For consumer products, and especially for anything young people use, the checks that run before and after a model call are part of the product, not overhead. We do not remove them to save money. See Safety hooks for consumer AI products. Paid outputs get the stronger model when tests say so. When someone pays for a personal report — a KeşifAtlası eligibility report, for example — the explanation is the product. If evaluation shows a stronger model explains complex rule combinations more clearly, the margin on the report can usually absorb it. Validation retries are worth paying for, up to a point. A retry that turns an invalid answer into a valid one is cheaper than a support ticket or a wrong answer. A retry loop that runs five times is a signal to fix the prompt, not to keep paying. Evaluation runs are an investment. Running a feature's test set against a new model or prompt costs money, and it is the cheapest insurance we buy. The test we apply is the one from the feature review above: does the extra spend produce value a user or the business would recognise? If yes, we keep it and budget for it. If a feature cannot justify its cost even after optimisation, that becomes an input to the continue, stop or scale decision rather than a problem to hide. Summary AI makes it possible for small teams to build powerful products. It also makes it possible for small teams to run up large bills without noticing. At Oryvelon, cost discipline is part of how every AI feature is designed: separate projects and budgets per product, the right model for each task, caching and batching where they are safe, output limits, monitoring for silent growth and a quarterly review of everything we pay for. The goal is simple: every company should know what its AI costs, what it earns from it, and whether the trade is worth making. For the broader principles behind our AI work, see AI explains, verified data decides. ## Company builder, venture studio or holding company? A practical comparison URL: https://oryvelon.com/insights/company-builder-vs-venture-studio-vs-holding/ Category: Company Building Ask ten people what a "company builder" is and you will get ten answers. Some will describe an agency with ambitions. Some will describe an investment fund. Some will say "a startup studio", as if the terms were interchangeable. The confusion is understandable — the models overlap — but it matters, because each one creates very different incentives. This is a practical comparison of the five models most often confused with each other: company builder, venture studio, holding company, agency and venture capital fund. For each we look at who builds, who owns, who operates, how money is made and when the model makes sense. At the end we explain why Oryvelon chose to be a company builder, and what that means for the people we work with. Three questions that separate the models Most of the confusion disappears if you ask three questions about any organisation: Who builds the product? The organisation's own people, external founders, or a client? Who owns the result? The organisation, a founding team, a client, or a mix? Who operates it after launch? Does the organisation stay involved in running the business, or does it step back? Here is how the five models answer: Model Who builds Who owns Who operates after launch Main source of value Company builder The builder's own team The builder The builder, with each company's team Value of the companies it owns Venture studio Studio team, then external founders Shared, founders increasingly Founders, with studio support early on Equity in spun-out companies Holding company Usually someone else The holding company Management of each business Dividends and asset value Agency Agency team The client The client Fees and retainers Venture capital fund Founders Founders and investors Founders Investment returns The table simplifies, but the pattern is clear. Only the company builder answers "us" to all three questions. The company builder A company builder creates companies internally, owns them and keeps operating them. Its advantage comes from reusing foundations across companies: infrastructure, security, measurement, AI governance, operating standards and the experience of people who have launched before. How it earns: through the value of the companies it runs — their revenue, profit and, eventually, their worth as businesses. What it controls: a lot. Strategy, product direction, data standards, capital allocation. Where it is strong: when several businesses share a meaningful technical or operational foundation, when the builder has relevant experience across them, and when it is prepared to make hard decisions about where to invest. Where it is weak: it can spread itself too thin, launch too much, or turn its shared layer into bureaucracy. It also concentrates risk: if the builder makes a poor standard decision, every company inherits it. We describe the model in detail in What is a digital company builder?. The venture studio A venture studio also creates companies from scratch, often from ideas generated internally. The difference is usually what happens next. Studios commonly bring in external founders or CEOs to lead each company, raise outside investment early and gradually reduce their own role. The studio keeps equity but gives up day-to-day control. How it earns: through the equity it retains in the companies it spins out. Where it is strong: attracting entrepreneurial talent who want to run their own company with a head start; moving quickly to external funding. Where it is weak: alignment can get complicated as founders, investors and the studio each hold different stakes and priorities. Standards set by the studio may not survive once the company becomes independent. The line between a studio and a company builder is blurry, and many organisations sit somewhere in between. The practical question is how long the parent stays in control and how much it continues to operate. The holding company A holding company owns businesses — sometimes fully, sometimes partially — and usually leaves their management to separate teams. Some holding companies are active and hands-on; many are primarily owners that allocate capital and appoint leaders. How it earns: dividends, profits and the value of its holdings. Where it is strong: long-term ownership, diversification, capital allocation across established businesses. Where it is weak: it typically does not build new companies itself and may have little operational influence over how each business runs. In legal terms, a company builder may well use a holding structure. The difference is not the legal form; it is the activity. A company builder builds and operates. A holding company, in the everyday sense, owns. The agency An agency builds for clients. A digital agency might design websites, build online stores, run advertising or manage social media. It earns fees for its work, and the client owns what is produced. How it earns: project fees and retainers. Where it is strong: hands-on execution across many clients; deep practical knowledge of what works in the market; steady revenue from ongoing services. Where it is weak: agency revenue is tied to hours and headcount; the agency does not own the value it creates for clients. This distinction matters to us more than any other, because it is the one most likely to be confused in our case. Oryvelon's founders have deep agency and e-commerce experience, and the group includes an agency — WeAreMedia. But Oryvelon itself is not an agency. It does not sell services to clients. Anyone looking for web, Shopify, advertising or SEO work should go to WeAreMedia. We explain why a group benefits from a separate services company in Why a group needs a dedicated services company. The venture capital fund A venture fund invests money in companies founded and run by others, in exchange for equity. It may offer advice, networks and support, but it does not build products. How it earns: returns when its investments grow in value and are sold or listed. Where it is strong: backing many companies, providing capital at scale, spreading risk across a portfolio. Where it is weak: limited operational influence; outcomes depend heavily on founders it did not choose to be. A company builder can work with venture funds — for example, if one of its companies raises capital — but it is a different kind of organisation. Hybrid models are common Real organisations rarely fit one box. An agency may launch its own products. A studio may keep some companies fully owned. A holding company may build a new business from scratch. A company builder may bring in a partner for one of its companies or raise outside capital for another. That is fine. The useful question is not "which label applies?" but "what does this organisation actually do, and what does that mean for me?" If you are considering working with one — as a founder, partner, employee or investor — ask about the three axes directly: who builds, who owns, who operates. How incentives differ The models create different incentives, and incentives shape behaviour more than labels do. An agency is rewarded for work delivered. It does well when clients need more work. The best agencies are excellent partners; the incentive, though, is to sell services, not to build long-term products of their own. A fund is rewarded for outliers. A few very large outcomes pay for many failures. That encourages bold bets and fast growth, sometimes at the expense of steady, profitable businesses. A studio is rewarded for successful spin-outs. It does well when companies raise money and grow quickly under external leadership. A holding company is rewarded for durable value. It tends to favour stable, cash-generating businesses. A company builder is rewarded for the combined value of the companies it runs. That encourages two things at once: building companies that are genuinely good businesses, and making the shared foundations better so every company benefits. It also creates a responsibility to stop companies that are not working, because weak companies consume attention the strong ones need. When each model makes sense There is no universally best model. Each fits a different situation. Choose an agency if your strength is execution for others and you want revenue from day one. Choose a fund if your strength is capital and judgement about founders, and you want broad exposure. Choose a studio if you want to generate ideas and attract entrepreneurs to lead them, with outside capital early. Choose a holding company if you want to own established businesses for the long term. Choose a company builder if you intend to build several businesses yourself, they share a real foundation, and you are willing to stay in the operating seat. Why Oryvelon is a company builder We chose the company builder model for four reasons. We want to own what we build. Years of building stores, advertising operations and digital products for others gave us a clear view of what it takes to make a digital business work. With Oryvelon, the value of that work stays with the companies we run. Our companies share a real foundation. MerchNivo, CastLyra, EduRelia, ZodiVela, KeşifAtlası, Noveniq, WeAreMedia and Sinem Keser Beauty Academy look very different to their customers. Underneath, they need the same things: secure domains and email, reliable deployment, measurement, AI governance, cost control. Building that foundation once and reusing it is where the model earns its keep. We want to stay in the operating seat. A studio's early handover to external founders would not suit how we work. We want to set the standards, watch the numbers and make the allocation decisions ourselves. We want services to stay separate. Keeping WeAreMedia as its own company lets client work have its own team, economics and relationships, while Oryvelon focuses entirely on building and operating companies. It also keeps the message clear: if you need services, there is one place to go. What this means if you work with us If you are a potential partner, you are dealing with the owner and operator of the company in question, not an intermediary. If you are a customer of one of our companies, the company you use is backed by a group that maintains its security, infrastructure and data standards — and keeps your data within that company. If you are considering investment in one of our companies in future, each company is designed to be separable, with its own code, data, infrastructure and economics. See Designing every company so it could stand alone. If you are looking for services, please talk to WeAreMedia. Oryvelon does not sell agency work. A quick self-test If you are trying to classify an organisation — including your own — these questions usually settle it: Does it have paying clients for whom it builds things they own? It has an agency component. Does it write cheques for equity in companies others founded? It has a fund component. Does it create companies and hand them to external founders? It is behaving like a studio. Does it own businesses without building or running them? It is behaving like a holding company. Does it build companies itself, own them and keep running them? It is a company builder. Oryvelon answers yes to the fifth question and no to the others. The only exception is inside the group: WeAreMedia is an agency, and it is run as a separate company for exactly that reason. A short glossary of related terms Several other terms appear in conversations about these models. Here is how we use them. Venture builder. Usually a synonym for company builder or venture studio. We avoid it because it blurs the difference between keeping operating control and handing it over. Incubator. A programme that supports early-stage founders, often with space, mentoring and small amounts of capital, for a fixed period. The founders build and own their companies; the incubator helps. Accelerator. Similar to an incubator but usually shorter and more structured, often ending with a demo day in front of investors. Again, founders build and own. Corporate venture. A large company's investment arm or internal new-business unit. It may build or invest, but its companies are usually tied to the parent's strategy. Operating company. The business that actually serves customers and earns revenue — in our case, each company on the Companies page. The group's role is to build and support operating companies, not to replace them. Portfolio. The set of companies a builder, studio, holding company or fund is involved with. In a company builder, the portfolio is actively managed: companies receive more or less attention based on evidence. See Capital allocation inside a company builder. Shared services. Functions provided centrally to several companies — for example security, infrastructure or finance. In a company builder, shared services should make companies faster, not add approval layers. Knowing these terms helps cut through marketing language. Whatever an organisation calls itself, the three questions — who builds, who owns, who operates — will tell you what it really is. One idea, five models: a worked example The differences become clearer when you run a single idea through each model. Take a problem we know well: independent Shopify merchants juggle many screens, and nobody watches the whole store every day. An agency would solve it store by store. A merchant hires the agency, which builds dashboards, sets up reports and perhaps sends a weekly review. The client owns the setup and pays for the hours. The agency learns a lot, but the value it creates stays with each client. A fund would wait for a founder who is already building a solution, assess the team and the market, and invest in exchange for equity. The fund would not write the product or decide what it measures. A venture studio might build a first version with its own team, then recruit a chief executive, raise outside money and step back to a board seat within a year or two. A holding company might look for an existing store-analytics business that is already profitable and buy it, keeping the management in place. A company builder builds the product with its own team, owns it and keeps running it. That is the path we took with MerchNivo. It sits on the group's shared foundations — the AI gateway, security standards, measurement — while its data, keys and customers stay its own. It also benefits from something no other model offers as directly: Noveniq, a store the group operates itself, where commerce assumptions can be checked against real operations. None of the five paths is wrong. They produce different companies, owned by different people, with different pressures on them. The trade-offs we accept Choosing the company builder model has costs, and we would rather name them than pretend they do not exist. Attention is the scarce resource. A studio can add companies by adding founders. We add companies only when we can run them properly, which limits how many we start. That is why most ideas stop inside our 10-step protocol, and why every live company faces a dated continue, stop or scale decision. Shared standards cut both ways. A good standard helps every company; a poor one harms every company. We review the shared layer on a fixed operating cadence so mistakes are caught before they spread. Sharing must stop at infrastructure. The temptation in a group is to join things up: one login, one customer database, one app. We keep shared infrastructure but separate data, and we explain why we resist merging in Why we don't merge our products into one app. Brands must earn trust on their own. A house of brands means each company builds its own reputation, with only a light "An Oryvelon company" endorsement. It is slower than one master brand, but it keeps each company credible in its market and separable later. See Naming, domains and brand architecture. There is nobody else to blame. When a builder owns and operates, the results are ours. We consider that a feature of the model. Summary The labels overlap, but the models do not. A company builder builds, owns and operates; a studio builds and hands over; a holding company owns; an agency builds for others; a fund invests in others. Each has its place. Oryvelon is a company builder because that is the model that fits what we want to do: build good companies on shared foundations and keep running them well. The companies we operate are on the Companies page, and how we build them is described on the Build page. ## How we validate a digital business idea before building it URL: https://oryvelon.com/insights/how-to-validate-a-digital-business-idea/ Category: Company Building The most expensive way to find out that nobody wants your product is to build it first. Yet that is how many digital businesses are started: someone has an idea, falls in love with it, spends months building, launches — and discovers that the people they imagined do not exist, will not pay, or cannot be reached at a sensible cost. At Oryvelon, Validate is the second stage of our operating model, after Discover and before Build. It is the cheapest stage in a company's life and the one with the most leverage. This note explains how we approach it. What validation is trying to answer Validation is not about proving that an idea is good. It is about finding out, as cheaply as possible, whether three things are true. The problem is real. A specific group of people experiences it often enough, and painfully enough, to want it solved. Someone will pay. Not "would find it useful" — will actually pay, at a price that can sustain a business. They can be reached. There is a realistic, affordable way to put the product in front of those people repeatedly. If any of the three is false, the idea does not become a company in its current form. It might change shape; it might wait for a better moment; it might stop entirely. All three outcomes are fine. Behaviour beats opinion The most important principle in validation is simple: trust what people do, not what they say. People are kind. Show them an idea and many will say it sounds great. Ask in a survey whether they would use it and a surprising number will say yes. Neither tells you much. The signals that matter are the ones that cost the person something — time, attention, money or reputation. Weak signal Strong signal "That's a great idea" Joining a waitlist with a real email Survey says "likely to use" Replying to a follow-up and booking a call Likes on a social post Paying a deposit or pre-ordering A friend's enthusiasm A stranger recommending it to someone else Interest in a demo Using a manual version more than once We design validation so that people have the chance to show strong signals. Step one: talk to people with the problem Before building anything, we talk to people who have the problem. Not about our idea — about their situation. For an idea like MerchNivo, that meant conversations with store owners about how they actually run their day: which screens they open, what they check, what they forget, what they would hand to someone else if they could. For KeşifAtlası, it meant understanding how people currently research visa options, where they get stuck and what they end up paying for. Good questions are about the past, not the future: "When did this last happen to you?" "What did you do about it?" "What did that cost you — in time or money?" "What have you tried already?" If people struggle to remember the last time the problem occurred, it probably is not painful enough. If they have already tried to solve it — with spreadsheets, freelancers, workarounds — the problem is real. Step two: write the offer before the product Next we write the offer as if the product existed: a clear description of what it does, who it is for and what it costs. This is harder than it sounds, and the difficulty is informative. If we cannot describe the value in a few sentences a customer would understand, the idea is still too vague. The offer becomes the core of the phase-zero site: a real page with real content explaining the problem and the solution, and a way to register interest. Step three: put the offer in front of real people A phase-zero site lets us measure how strangers respond to the offer. We look at: how many visitors understand it well enough to act (join, reply, request access); which messages and headlines produce more action; where visitors come from, and which channels bring people who act; what questions people ask when they get in touch. Because the site is built properly — clear structure, fast pages, sensible search markup — it also starts building discoverability long before launch. See Search and AI-assistant visibility. Step four: run a manual version One of the most useful validation techniques is to deliver the value manually before automating it. Sometimes called a "concierge" test, it means doing by hand what the product will eventually do by software. For a store-operations product, that might mean reviewing a handful of stores' data manually and sending each owner a written weekly briefing. For an eligibility product, it might mean evaluating a small number of cases by hand against the rules and producing a report. Manual versions teach things that no survey can: what people actually value in the output; which parts they ignore; what they ask for that we had not thought of; whether they come back for more; how long the work really takes, which informs pricing and automation priorities. Step five: test the price Price is where many ideas quietly fail. We test it directly rather than assuming. The simplest test is to state a price and see what happens. On a phase-zero site, that might be a clearly stated price on the offer page. In conversations, it might be asking whether someone would pay a specific amount — and, better, whether they would pay now for early access. Reactions to a price are more useful than reactions to an idea. "That's too much" followed by an explanation tells you what value the person perceives. "Can I pay annually?" tells you something else entirely. A deposit, a pre-order or a paid pilot is the strongest validation signal there is. For subscription products we also test the unit: per store, per seat, per student, per report. The right unit often matters more than the exact number. We discuss pricing in more depth in Subscription plus usage: pricing logic for AI SaaS. Step six: check the channel The third question — can we reach these people affordably? — is often left until after launch, which is too late. During validation we try at least one or two acquisition channels at small scale: search, by publishing genuinely useful content on the phase-zero site; communities where the target users already gather; partnerships with businesses that already serve them; the group's own reach, such as founder authority sites and oryvelon.com, measured carefully with consistent UTM parameters; small paid tests, where the economics make sense. We are not looking for a perfect channel. We are looking for evidence that at least one channel can bring the right people at a cost the business can support. The signals we deliberately ignore Some signals feel meaningful and are not. Total visitors. A spike of traffic from a lucky post says little about whether the right people care. Compliments from friends and family. Kind, but not evidence. Competitor funding. A well-funded competitor suggests the problem may be real, but tells you nothing about whether you can win. Our own excitement. The founders being excited is necessary but nowhere near sufficient. Feature requests from people who would not pay. Useful for understanding the problem; dangerous as a roadmap. How validation fits the protocol Validation runs alongside the early steps of our 10-step protocol. Conversations inform the one-sentence problem. Pricing tests inform the revenue mechanism. Repeat behaviour in manual pilots informs the recurring loop. The phase-zero site is step eight in the protocol, but a simple version of it often appears during validation. By the time an idea reaches the Build stage, we expect to have: a clearly stated problem, confirmed by people who have it; an offer that strangers understand and act on; evidence that someone will pay, ideally money already committed; at least one channel that can bring the right people affordably; notes from a manual version showing what users actually value. If we cannot say those things, the idea is not ready to build. How long validation takes Less time than most people expect, and more time than most people spend. A focused validation phase is usually measured in weeks, not months. The goal is not certainty — there is no certainty in a new business — but enough evidence that building is a reasonable bet rather than a hopeful one. The risk to watch for is validation that never ends. If we keep running "one more test" because we are afraid of the answer, we are not validating any more; we are procrastinating. The continue, stop or scale discipline applies here too: at a pre-agreed point, we decide. When validation says no Many ideas fail validation. That is the point. A few common outcomes: The problem is real but nobody will pay enough. Sometimes this means a different customer segment will. Sometimes it means the idea belongs as a feature in another product rather than a company of its own. People will pay but cannot be reached affordably. This often calls for a partnership model or a different entry point. The problem is smaller than it looked. People manage well enough with what they have. The idea is right but the timing is wrong. A regulation, technology or market shift may change the picture later. We write down what would need to be true and revisit it. In every case, the work is not wasted. Conversations, offers and experiments become knowledge the group keeps. More than once, a failed validation for one idea has sharpened the understanding of a problem another company solves. Validation in a company builder versus a single startup Validating inside a group has a few advantages we try to use well, and a few risks we watch for. Advantages. The foundations already exist, so a phase-zero site with proper measurement can go live quickly. The group has experience across several markets, which helps with channel and pricing tests. And the founders' existing networks — e-commerce operators, brands, creators, educators — make it easier to find people to talk to. Risks. It is tempting to skip validation because launching is easy. It is also tempting to use the group's reach to create a flattering early signal — traffic from oryvelon.com is not the same as demand from strangers. We separate those sources in our measurement so we are not fooled by our own audience. A validation checklist For anyone validating an idea, here is the short list we use. [ ] We have spoken to people who have the problem, about their past behaviour. [ ] We can describe the offer in a few sentences a customer understands. [ ] A real page with that offer is live and measured. [ ] We have delivered the value manually to at least a few people. [ ] We have stated a price and seen real reactions — ideally payments. [ ] At least one channel has brought the right people at a sensible cost. [ ] We have written down what we learned and what would make us stop. If most boxes are ticked, the idea is ready for the Build stage. If not, the missing boxes are the work. Conversation guides by business type The questions we ask change with the type of business. Here are the guides we start from. For software sold to businesses (the MerchNivo type): - Walk me through the last time you dealt with this. Which tools did you open? - Who else in the business touches this work? - What do you pay today — in software, freelancers or staff time — to handle it? - If this disappeared tomorrow, what would break first? For marketplaces (the CastLyra type), we run two guides, one per side: - Businesses: how did you find the last person you hired for this? How long did it take? What went wrong? - Talent: where do your best opportunities come from? How do you tell a serious request from a waste of time? For consumer products (the ZodiVela type): - When did you last use something like this? What made you open it? - What did you like about it, and what made you stop? - Have you ever paid for something similar? What made it worth it? For decision-support products (the KeşifAtlası type): - Where did you look for information? Which sources did you trust? - What decision were you trying to make, and what was at stake? - Did you pay anyone for advice? What did you get for it? For education and services (the EduRelia and Sinem Keser Beauty Academy types): - Who decides to buy, and who uses it? Are they the same person? - What does progress look like to you, and how do you know it happened? - What would make you recommend this to someone else? The guides are starting points, not scripts. The best conversations follow the person's story. Write the stop criteria before the test The most useful habit in validation takes five minutes: before running any experiment, write down what result would make you stop, change course or continue. Once the results arrive, it is too late. Every number can be explained away by someone who wants the idea to live. We keep the format simple: Question Continue if… Change shape if… Stop if… Is the problem real? Most people we speak to describe a recent instance and a workaround The pain is real but belongs to a different group than we assumed People struggle to recall the problem at all Will someone pay? Some people commit money or a firm next step People pay for part of the offer, not the whole Interest disappears as soon as a price appears Can we reach them? One channel brings the right people at a cost the model supports Only a partner could reach them affordably No channel we can afford brings people who act The thresholds behind "most" and "some" are set per idea, in writing, and agreed before the test starts. A narrow B2B product may need only a handful of committed customers to justify building; a consumer product needs evidence at a larger scale. What matters is that the line is drawn first. It also helps to estimate, roughly, whether the business can work at the price being tested. A few lines on what it will cost to serve one customer — hosting, AI usage, support time, payment fees — often reveal that a promising price leaves no margin. We cover that arithmetic in Unit economics per product. Validation experiments by business shape The conversation guides above are the same for every idea in spirit. The experiments that follow them differ, because different kinds of business fail in different ways. Marketplaces must validate both sides, and the order matters. For a business like CastLyra, talent interest without business demand is a directory, and business demand without verified talent is a disappointment. Validation asks which side is harder to attract and whether the first participants on that side can be recruited by hand. See Solving the cold-start problem. Decision-support products can let the free step do the validating. For KeşifAtlası, a short free eligibility test is both a useful tool and a validation instrument: it shows how many people complete it, which situations are most common and how many ask for more detail. The step from that test to a paid report is the price test. We describe the design in From free test to paid report. Commerce brands validate with small, real orders. For a product brand like Noveniq, the honest test is a limited run of real products, sold to real customers, with every cost counted — including returns and delivery problems. A store that runs for real also teaches lessons that feed the group's software. See Using a real store as a testbed. Consumer subscriptions validate retention, not sign-ups. Signing up for a free reading or a trial is easy. The signal that matters for a product like ZodiVela is whether people come back on their own the following week, and whether some choose to pay for more. In each case the experiment is chosen to test the most likely reason this particular kind of business would fail. That is the whole idea: find the weak point cheaply, before it is built into the product. Summary Validation is where good companies save the most money and where weak ideas should end quietly. At Oryvelon we look for behaviour over opinion, test offers and prices before code, deliver value manually before automating it, and check that customers can be reached before committing to build. It is not glamorous work, but it is the reason the companies on our Companies page are companies rather than experiments. ## Phase-zero sites: launching a brand before the product exists URL: https://oryvelon.com/insights/phase-zero-sites/ Category: Company Building There is a familiar sight on the internet: a domain with a logo, the words "coming soon" and an email field. It tells visitors almost nothing, gives search engines nothing to index and teaches the founders nothing about whether their message works. At Oryvelon we do something different. Every company gets a phase-zero site — a real website with real content — before its product is finished. It is step eight of our new-company protocol, and in our experience it is one of the highest-leverage things a new company can do. What a phase-zero site is A phase-zero site is the first real version of a company's website. It exists before the product launches, and it does four jobs: Explains the problem in the customer's own language. Describes the solution clearly enough that the right person recognises it is for them. Captures interest through a waitlist, an early-access request or a direct contact form. Starts earning discoverability in search engines and AI assistants. It is not a placeholder. It is not a single landing page with a headline and a form. It is a small, well-built site with enough substance to be useful — and to be worth indexing. What it must contain The exact pages vary by company, but a phase-zero site usually includes: A clear homepage that states who the product is for, what problem it solves and what makes it different. A "how it works" page describing the approach, even if some parts are still being built. Pages for each audience, where there is more than one — a marketplace needs to speak to both sides; an education product may need separate messages for families and schools. Answers to real questions, often as an FAQ built from early customer conversations. A way to act: join a waitlist, request early access, book a call or get in touch. Honest status information so visitors know what is available now and what is coming. Legal basics: privacy notice and terms appropriate to what the site collects. The group relationship, stated naturally — usually a small "An Oryvelon company" in the footer and a line on the about page. What it must not do A phase-zero site must never pretend the product is finished when it is not, promise features the team is not confident about, or publish invented testimonials, metrics or customer logos. Credibility is the whole point of the exercise, and it is easy to destroy. It also must not collect more data than it needs. A waitlist needs an email address; it probably does not need a phone number, a birthday or a company size. The same minimum-collection rule that applies to products applies here. Built properly from day one Because the phase-zero site is the company's public face, it is built on the same foundations as a live product: Domain security: registered, locked, set to auto-renew, under an account with two-factor authentication. HTTPS and security headers: the same standards as every Oryvelon site. Email authentication: SPF, DKIM and DMARC on the domain, so waitlist confirmations arrive and nobody can easily spoof the brand. Measurement: analytics that respects visitors, with events for the actions that matter — sign-ups, replies, requests. Search structure: clean headings, descriptive titles and meta descriptions, a sitemap, sensible robots rules and structured data describing the organisation. AI readability: clear, factual content that assistants can understand and quote, plus machine-readable summaries where appropriate. We cover the discoverability side in Search and AI-assistant visibility and Structured data and llms.txt. Why launch the site before the product? It tests the message with strangers Founders explain their idea to friends, colleagues and investors, who usually understand it because they know the context. Strangers do not. A phase-zero site shows whether someone arriving cold understands the offer well enough to act. If they do not, it is far cheaper to change a headline than a product. It builds an audience early Every person who joins a waitlist before launch is someone who can be invited to try the first version, give feedback and become an early customer. Launching to a list of interested people is very different from launching to silence. It starts search visibility months ahead Search engines take time to discover, index and trust a new site. A company that launches its website on the same day as its product starts that clock late. A company whose phase-zero site has been live and useful for months arrives at launch with pages already indexed and, often, early rankings for relevant questions. The same applies to AI assistants. When people ask an assistant about a problem, assistants draw on content they can find and trust. A clear, factual phase-zero site is part of how a new company becomes something an assistant can mention. It gives the brand a home Partners, journalists and potential hires will look a company up. A proper site answers their questions and signals seriousness. A "coming soon" page signals the opposite. How we measure a phase-zero site We measure a phase-zero site against the questions validation is trying to answer, not against vanity numbers. Question What we look at Do visitors understand the offer? Share of relevant visitors who take a meaningful action Which message works best? Action rates across headline and page variants Who is interested? Answers to one or two optional questions at sign-up Where do interested people come from? Sources and campaigns of visitors who act, not all visitors Is interest real? Replies to follow-up emails, calls booked, early access accepted Is search starting to work? Indexed pages, impressions and queries in search tools Traffic from the group's own channels — oryvelon.com, founder sites, sister brands — is tagged with consistent UTM parameters so it can be separated from organic interest. A spike of visits from our own audience is encouraging, but it is not validation. See UTM standards for measuring traffic between sister brands. Content that earns its place The content on a phase-zero site should be useful even to a visitor who never signs up. That is what makes it worth indexing and worth citing. Good phase-zero content often comes straight from validation conversations: Explanations of the problem in the words customers use. Guides that help people with part of the problem today, before the product exists. Honest comparisons of the ways people currently handle the problem. FAQs built from real questions asked in conversations and sign-up forms. For a relocation product like KeşifAtlası, that might mean clear explanations of how eligibility generally works and what documents people usually need to think about — carefully written, with the product's rules engine as the source for anything specific. For a store-operations product like MerchNivo, it might mean practical guides on the signals a Shopify store should watch every day. See The operational signals every Shopify store should watch. What we avoid is thin content written only to rank. It does not help visitors, it weakens the brand and it is exactly the kind of content search engines are designed to ignore. From phase zero to launch A phase-zero site is not thrown away at launch. It evolves. The waitlist form becomes a sign-up or trial flow. The "how it works" page gains real screenshots and details. The FAQ grows with questions from real users. Status notes are updated as features arrive. Content pages keep earning traffic and become part of the company's long-term marketing. Because the site was built on the group's standards from the start — security, measurement, search structure — there is no migration or rebuild at launch. The company simply turns the product on. Phase-zero sites across the group Each Oryvelon company has a public home on its own domain, and each is listed on the Companies page with a profile describing what it does. The profiles on oryvelon.com complement the companies' own sites: they explain each company in the context of the group, while the company site speaks directly to its customers. The stage of each site varies. Some are fully live products; some are earlier in their journey. The standard is the same: real content, honest status, proper foundations. Common mistakes A few mistakes come up often when teams build pre-launch sites: Only a "coming soon" page. It wastes months of potential search visibility and teaches nothing. Over-promising. Listing features that are not certain creates disappointment and support work later. Invented social proof. Fake testimonials or logos destroy trust instantly when discovered. No measurement. Without events on the actions that matter, the site cannot validate anything. Collecting too much. Long sign-up forms reduce sign-ups and create data you must protect. Forgetting email authentication. Waitlist emails land in spam, and the domain becomes easier to spoof. Treating it as temporary. A site built to be thrown away rarely gets the care that makes it valuable. A phase-zero checklist Before a phase-zero site goes live at Oryvelon, we check: [ ] Domain locked, auto-renewing, under a 2FA-protected account. [ ] HTTPS, security headers and email authentication in place. [ ] Homepage states the audience, problem and solution clearly. [ ] Each audience has a page that speaks to it. [ ] At least one useful piece of content beyond the pitch. [ ] A clear action, collecting only necessary data. [ ] Honest status information. [ ] Privacy notice and terms. [ ] Analytics with events for key actions; internal traffic tagged. [ ] Titles, descriptions, sitemap, robots rules and organisation structured data. [ ] "An Oryvelon company" mentioned naturally, not as a logo. Writing the homepage: a practical template The homepage carries most of the weight on a phase-zero site. We use a simple structure that works for very different companies. 1. One sentence that names the person and the outcome. Not the technology, not the company's ambition — the person and what changes for them. "Your Shopify store, watched every day" says more than "AI-powered commerce intelligence". 2. The problem, described the way customers describe it. Two or three short paragraphs using phrases from real conversations. Visitors should recognise their own situation. 3. How the product helps. Three or four concrete capabilities, each tied to part of the problem. If a capability is not ready yet, say so. 4. Who it is for — and who it is not for. Being explicit about fit saves everyone time and makes the right visitors more confident. 5. What happens next. A single clear action and a sentence about what the visitor will receive: an invitation, an early-access email, a call. 6. Trust signals that are true. The team behind the product, the group it belongs to, how data is handled. No invented logos or quotes. 7. Answers to the questions people actually ask. Pulled from conversations, not imagined. This structure is deliberately plain. The design can be distinctive; the message should be unmistakable. Waitlist emails that people read A waitlist is only useful if the relationship continues after the sign-up. We keep waitlist communication short, honest and infrequent. Confirmation: sent immediately, from the company's own domain with proper authentication, confirming what the person signed up for and what to expect. Occasional updates: real progress, not filler — a feature that now works, a question we want their opinion on, an invitation to try something. The invitation: when early access opens, a clear message with a single action. An easy exit: every email makes it simple to leave the list. Each company sends its own emails to its own list. A waitlist for one Oryvelon company is never used to promote another. See Marketing consent is not product consent. When phase zero ends There is no fixed duration. A phase-zero site moves into launch when the product meets its written MVP acceptance criteria and the team is ready to support real users. Before that point, the site keeps doing its job: learning which messages work, building the list, and accumulating search visibility. Occasionally the phase-zero site reveals that the idea needs to change — a different audience responds, or a secondary feature turns out to be the main attraction. That is a success, not a setback. It is exactly the kind of lesson that is cheap to learn at phase zero and expensive to learn after launch. Phase-zero sites and the Oryvelon website Oryvelon.com itself follows the same idea at group level. It explains what the group is, lists the companies, describes how we build and publishes notes like this one. Each company profile on oryvelon.com — for example MerchNivo, CastLyra or EduRelia — acts as a second, independent explanation of that company, written in the context of the group. Together with each company's own site, that gives search engines and AI assistants two consistent, factual sources describing what the company does and who stands behind it. That consistency matters. When a company's own site, its profile on oryvelon.com and its founders' authority sites all describe it in the same terms, it becomes much easier for search engines and assistants to understand the company as a real entity — and to recommend it when someone asks the right question. See Founder authority sites and how they relate to a company group. Phase-zero sites for different kinds of company The template above works across the group, but each type of company has its own emphasis, and its own traps. A marketplace needs two front doors. A site for CastLyra has to speak to talent and to businesses, with separate pages, separate calls to action and separate sign-up forms. Mixing them confuses both. It also must not imply activity that does not yet exist: no talk of a thriving community before there is one. Registration forms ask only for what verification will later need, and nothing is published about any person without their consent. An education product speaks to adults. For EduRelia, the phase-zero site addresses families and schools, and the waitlist collects contact details from adults only. There is no reason to collect anything about a child before the product exists, so the site does not. A consumer AI product sets expectations early. A site for ZodiVela states plainly that readings are for reflection and entertainment, not prediction or advice. Setting that framing at phase zero is far easier than correcting it after launch. A decision-support product is careful with specifics. A relocation site can explain in general terms how eligibility tends to work, but specific answers belong to the product's verified rules, not to a blog post that may go out of date. Where rules change, the content says when it was last reviewed. A commerce brand can sell from day one. For a direct-to-consumer brand, phase zero may simply be a small catalogue with honest stock information and a notify-me option for items not yet available. The same principles apply: true claims, minimal data, proper measurement. How much to build at phase zero There is a real trade-off here. Build too little and the site teaches nothing and ranks for nothing. Build too much and you have spent product time on a website for an offer that may still change. The rule we use: build what you would keep, and nothing you would have to undo. That means the foundations are done properly — domain, DNS and email authentication (see Domain, DNS and email security), measurement, search structure — because they carry straight into launch. It also means content that answers real questions, since it keeps earning after launch. It usually does not mean user accounts, complex content management, a design system built out for every future screen, or pages for features that are still uncertain. Those can wait until the product exists and the message has been tested. A few specific choices follow from that rule: Measure actions, not people. Phase zero needs to know how many visitors joined, from which source and after reading which page. It does not need to follow individuals around. See Analytics without surveillance. Settle the name and domain first. Rebranding after months of indexing throws away the search visibility the site was built to earn. See Naming, domains and brand architecture. Keep pages easy to change. Headlines and offers will be rewritten several times. A simple, well-structured site makes that a small job. A good phase-zero site is modest in scope and high in quality. That combination is what lets it turn into the launch site without a rebuild. Summary A phase-zero site is the first real product a company ships: a clear, honest, well-built website that tests the message, builds an audience and starts earning discoverability before the product is ready. It costs little, teaches a lot and keeps paying back long after launch. That is why every Oryvelon company starts with one. ## Continue, stop or scale: how we make portfolio decisions at 30, 60 and 90 days URL: https://oryvelon.com/insights/continue-stop-or-scale/ Category: Company Building Launching a company is exciting. Deciding what to do with it afterwards is harder — and far more important. Without a clear decision process, new companies tend to drift. Nobody wants to be the person who says "this isn't working", so the company keeps running on hope. Months pass. Attention and money that could have gone to something stronger are spent on something that stopped improving long ago. At Oryvelon every company — and every major new initiative within a company — has written KPIs for 30, 60 and 90 days after launch, and a commitment to make one of three explicit decisions at each checkpoint: continue, stop or scale. This note explains how that works. Write the targets before launch The most important rule is about timing: KPIs are written before launch, not after. After launch, it is very easy to look at whatever happened and find a way to call it a success. "We didn't hit our sign-up target, but engagement is great." "Revenue is low, but we learned a lot." Sometimes those statements are true. Often they are a way of avoiding a decision. Writing targets in advance — as step ten of our new-company protocol — removes that temptation. The targets can be wrong; that is fine. What matters is that they were stated honestly beforehand and that any change to them is a deliberate, recorded decision. Choose KPIs that measure behaviour Good checkpoint KPIs measure what customers do, not what we do or what they say. Stage Weak KPI Strong KPI 30 days Pages published, visitors Share of sign-ups who complete a first meaningful action 60 days Features shipped Share of active users still active after several weeks 90 days Total revenue in isolation Paying customers, retention and gross margin per customer The specific numbers differ by company. For MerchNivo, activation might mean a store connected and a first briefing opened; for CastLyra, verified profiles on one side and requests sent on the other; for KeşifAtlası, eligibility tests completed and reports purchased; for Noveniq, repeat purchases. We try to keep the list short — three to five KPIs per checkpoint. A long list makes it easy to find one metric that looks good and ignore the rest. The three checkpoints Day 30: is anything working? Thirty days is too early to judge a business, but it is not too early to see whether the basics work. The day-30 checkpoint asks: Are the right people arriving? Do they understand what the product does? Are they completing the first meaningful action? Is anything technically broken or unexpectedly expensive? Most day-30 decisions are continue, often with a list of fixes. A day-30 stop is rare and usually means something fundamental was wrong — the audience is not who we thought, or the product cannot deliver its core promise. Day 60: are people coming back? By sixty days, the question shifts from first use to repeated use: Do activated users return? Are early paying customers staying? Is the cost of serving each customer in line with expectations? Is at least one acquisition channel showing sustainable economics? This is often the most revealing checkpoint. A product can generate curiosity for a month on novelty alone. Return behaviour at day sixty says whether the product solves a real, recurring problem. Day 90: is this a business? At ninety days we ask whether the company is on track to become a business worth running: Is revenue growing from paying customers who stay? Does each customer generate more than they cost to acquire and serve, or is there a credible path to that? Is the team learning faster than the problems are growing? Compared with other opportunities in the group, does this deserve more, the same or less attention? The day-90 decision is the one most likely to be scale or stop. What each decision means The three decisions are deliberately simple. Each has a clear meaning. Continue means the evidence is good enough to keep going on the current plan, with any adjustments written down. It is not a default. A continue decision should come with a short statement of why, and what we expect to see at the next checkpoint. Stop means the company or initiative ends in its current form. That might mean shutting it down, pausing it until conditions change, folding its useful parts into another company, or reshaping it into something different that starts the protocol again. Stopping is described more below, because it is the hardest decision to make well. Scale means the evidence is strong enough to put more resources behind it: more budget for acquisition, more engineering time, more of the founders' attention, perhaps outside partners. A scale decision should also specify what must stay true — margins, retention, quality — as the company grows. Evidence we weight heavily At each checkpoint we look at a set of evidence, and some of it counts more than the rest. Heavily weighted: - activation and retention curves; - paying customers and whether they renew or repurchase; - cost to serve each customer, including AI costs; - qualitative feedback from customers who use the product regularly; - whether a sustainable acquisition channel is emerging. Lightly weighted: - total visitors; - social media attention; - feedback from people who do not use the product; - the team's excitement — important for morale, weak as evidence. Ignored: - competitors' fundraising announcements; - metrics that were not defined before launch and appear only because they look good. Why stopping is a skill In most organisations, stopping something is treated as failure. In a company builder, it has to be treated as a skill — one we practise deliberately. Attention is finite. Every company in the group competes for the founders' time, engineering capacity and budget. A company that should have stopped takes those resources from one that should have scaled. Sunk cost is invisible but powerful. The more we have invested, the harder it feels to stop. Pre-agreed checkpoints make it easier to look at the evidence rather than the history. Stopping well preserves value. When a company stops, its domain, content, code, learnings and relationships remain. Useful parts can move to other companies. The knowledge flows back into the group's standards. A well-documented stop is often the start of something better. Stopping protects trust. A company kept alive without investment tends to decline slowly — slower support, fewer updates, frustrated customers. Stopping clearly, with proper notice and data export options for anyone affected, is more respectful than fading away. How we stop responsibly When the decision is to stop a live product, we follow a few rules: Tell customers early and clearly, with a reasonable timeline. Offer data export so customers leave with what belongs to them. See Retention, export and deletion. Refund fairly where customers have paid for service they will not receive. Delete what should be deleted according to the product's retention policy. Keep the domain and, where appropriate, a page explaining what happened. Write down the lessons, including what the checkpoints showed and when. Move useful assets — content, code, research — to where they can be reused. Checkpoints for features, not only companies The same discipline applies inside companies. A large new feature — an AI capability, a new pricing tier, a new market — gets its own short list of KPIs and checkpoints. Features that do not earn their keep are removed or simplified. This keeps products focused and prevents the slow accumulation of complexity that nobody uses. How checkpoints connect to capital allocation Checkpoint decisions are the raw material of the group's capital allocation. When several companies reach checkpoints, the decisions are compared: which companies deserve more investment, which should hold steady, which should stop. We write about that process in Capital allocation inside a company builder. Because each company's economics are tracked separately — revenue, hosting, database, AI, email and tools — the comparison is based on real numbers rather than impressions. See Unit economics per product. A checkpoint template Each checkpoint review follows the same short template, written by the company's owner and discussed with the group. KPIs as written before launch, with the actual numbers next to them. What surprised us, positively and negatively. What customers told us, quoting real words where possible. Cost to serve, broken down. The proposed decision — continue, stop or scale — and the reasoning. What we expect at the next checkpoint, if the decision is continue or scale. A review should fit on two pages. If it takes longer to explain, the situation is probably less clear than it should be — which is itself useful information. Common failure modes We watch for a few patterns that undermine checkpoint decisions: Moving the goalposts — changing KPIs after seeing the results. Averaging away the signal — reporting totals that hide weak retention. Permanent "continue" — continuing without stating why or what should change. Scaling too early — putting money into acquisition before retention is proven. Stopping too quietly — letting a product decline without a clear decision. Naming these patterns makes them easier to spot in ourselves. Examples of checkpoint reasoning Abstract rules are easier to apply with examples. The scenarios below are illustrative — simplified patterns of the kind every product team meets — rather than reports on specific results. A strong start that fades. A consumer product gets excellent sign-ups in its first month, driven by novelty and a well-received launch. At day 60, few of those users are returning. The right decision here is usually continue with a narrow focus: stop spending on acquisition, study the small group who do return, and rebuild the core experience around what they value. Scaling would multiply the leak. A slow start with loyal users. A business product signs up far fewer customers than planned in its first month. But nearly all of them are still active at day 60, and several have asked for features that would let them use it more. This is often a continue, then scale pattern: the product is solving a real problem for a narrow group, and the question becomes how to reach more people like them. Healthy usage, unhealthy economics. A product is used and loved, but the cost to serve each customer — perhaps driven by AI usage — is higher than what they pay. The decision is not stop or scale; it is continue with an economics fix: model routing, caching, pricing changes or usage-based tiers. See Subscription plus usage. A marketplace that fills one side only. Talent sign up enthusiastically; businesses do not. Scaling talent acquisition would make the imbalance worse. The decision is to continue with a single focus on the scarce side until requests start flowing. See Two-sided marketplaces: thinking about the cold-start problem. Nothing moves. Visitors arrive but do not act, even after several changes to the message and offer. Customers who try the product do not return. At day 90, the honest decision is usually stop, preserve what was learned and redirect the attention elsewhere. Who makes the decision Checkpoint decisions are made by the company's owner together with the group, not by committee. The owner proposes; the group challenges the evidence and the reasoning; a decision is recorded. If there is disagreement, the default is the option that preserves more future choices — usually continuing on a smaller budget until the next checkpoint rather than scaling or stopping irreversibly. The record matters. Six months later, when someone asks why a company was scaled or stopped, the answer should be a short document with the numbers and the reasoning, not a memory. What scaling actually involves "Scale" can sound like a single switch. In practice it is a set of specific commitments, and we write them down when the decision is made. Acquisition budget: how much more we will spend, on which channels, and what cost per customer we will accept. Product capacity: which improvements the extra engineering time will go to, prioritised by what retained customers value most. Operational readiness: support, onboarding and documentation that can handle more customers without quality dropping. Guardrails: the retention, margin and quality levels that must hold as volume grows. If they slip, scaling pauses. Next review: a date — usually sooner than the next normal checkpoint — to confirm that the extra investment is producing the expected results. Scaling a company that has not proven retention is one of the most common and expensive mistakes in digital business. Writing guardrails into the decision is how we avoid it. Setting targets when you have no history The first objection to writing KPIs before launch is obvious: how can you know what "good" looks like for a product nobody has used yet? You can't, precisely. But you can do much better than guessing. We build first targets from three sources: Validation evidence. The phase-zero site and early conversations tell us roughly how many interested people arrived, how many left an email or joined a waitlist, and what they said they would pay. That sets a floor for day-30 activation. The cost floor. Every company has a monthly cost it must eventually cover — hosting, database, AI usage, tools, a fair share of the core. Working backwards from that number tells us how many paying customers the day-90 target has to point towards, even if day 90 is too early to reach it. Patterns from similar products. Retention in a daily-use tool behaves differently from retention in a product people use once a season. We set expectations by product type, not by the best numbers we have read about elsewhere. We write the target as a range with a short sentence of reasoning — "we expect between X and Y because the waitlist converted at about this rate". When a target turns out to be badly wrong, the reasoning shows which assumption failed. That is often the most useful finding of the whole checkpoint. Worked example: a MerchNivo feature through three checkpoints Here is how the process looks for a single feature. The scenario is illustrative, not a report of results. Imagine MerchNivo adds an alert for unusual refund activity: when refunds in a store rise well above that store's own normal pattern, the daily briefing flags it, shows the orders involved and suggests what to check. The numbers come from the store's own data; the AI explains what changed and why it might matter. We describe the wider approach in Shopify operations signals. The KPIs written before release might be: Checkpoint Question KPI Day 30 Do merchants notice and understand the alert? Share of alerts opened; share followed by a suggested action Day 60 Is it useful enough to keep? Share of stores that keep the alert on; share of alerts merchants mark as "not useful" Day 90 Does it earn its place? Retention of stores that use it versus those that do not; cost per alert against plan price At day 30, suppose merchants open most alerts but rarely act. That is a continue with fixes — probably the suggested action is vague, or the alert fires on refunds the merchant already knew about. At day 60, suppose many merchants mark alerts as noise during a sale period, when refunds naturally rise. The fix is to compare against seasonal patterns, not a flat average. By day 90, the question is whether the feature changes behaviour that matters. If stores using it retain better and the cost per alert is small, it becomes a standard part of the briefing. If not, it is simplified or removed — without drama. Before the checkpoint: making sure the evidence is real A checkpoint is only as good as the data behind it. A week before each review, the company's owner works through a short checklist: [ ] Tracking for every KPI has been tested end to end, not assumed. [ ] Test accounts, staff activity and internal orders are excluded from the numbers. [ ] Cohorts are defined by sign-up date, so new arrivals do not flatter retention. [ ] Costs for the period are attributed to the company, including AI usage. [ ] At least a handful of real customer quotes have been collected from people who use the product. [ ] Any change to a KPI since launch is listed with the date and reason. We measure what the decision needs and nothing more; checkpoints do not require tracking individuals across the web. See Analytics without surveillance. The most common problem this checklist catches is not dishonesty. It is a broken event or an internal test account quietly inflating a number. Finding that the week before is far better than discovering it after a scale decision. Summary Every company we launch comes with a promise to ourselves: we will look honestly at the evidence at 30, 60 and 90 days, and we will make a clear decision. Continue, stop or scale. The discipline protects our customers, our team and the group's capital — and it is what makes the companies on our Companies page the ones worth running. ## Designing a recurring revenue loop from day one URL: https://oryvelon.com/insights/recurring-revenue-loops/ Category: Company Building Some businesses are like treadmills. Every month starts at zero: new customers must be found, convinced and sold to, and last month's customers are gone. That can work — many good businesses operate this way — but it is exhausting, expensive and fragile. A bad month in acquisition becomes a bad month in revenue immediately. Other businesses compound. Customers who pay this month are likely to pay next month. Existing customers bring new ones. Each month starts from a base that is higher than the last. At Oryvelon, step three of our new-company protocol asks a direct question: how does revenue repeat without constant manual effort? We ask it before the product is built, because the answer changes what we build. This note explains how we think about it. Three ingredients of a recurring loop Every recurring revenue loop we have seen work has three ingredients. 1. A recurring problem. The customer has the problem again and again — every day, every week, every term, every season. A problem that happens once rarely supports recurring revenue, no matter how clever the pricing. 2. A recurring moment of value. The product delivers something the customer notices, regularly. A daily briefing, a weekly match, a new reading, a term report, a restock. If value is delivered once and then becomes invisible, customers eventually ask why they are still paying. 3. A low-friction way to pay again. Renewal should be simpler than cancellation is tempting. That does not mean hiding the exit — it means making continued use effortless and fairly priced. If any ingredient is missing, recurring revenue tends to leak, and no amount of marketing will fix it for long. The loops we use across the group Oryvelon companies serve very different customers, so their loops differ. Seeing them side by side is useful. Company Recurring problem Recurring moment of value Revenue loop MerchNivo Running a store every day Daily priorities and operational alerts Subscription plus usage CastLyra Brands need talent continuously; talent need work continuously New matches, requests and bookings Business subscriptions, talent premium, later booking commission EduRelia Learning happens every week of the school year Progress, support and visibility for families and teachers Family subscriptions and annual school licences ZodiVela People return to reflection regularly New readings and personalised content Subscriptions, credits and one-off premium readings KeşifAtlası Circumstances and rules change; families plan in stages Updated eligibility and personal reports Paid reports, premium services and partner referrals Noveniq Devices change; accessories wear out New products for new devices Repeat purchases WeAreMedia Brands need ongoing digital execution Monthly campaigns, content and reports Retainers Sinem Keser Beauty Academy Beauty services recur; skills build over time Appointments, courses and new digital products Services, education programmes and digital products No single model is best. The right loop is the one that matches how often the customer has the problem and how value is delivered. Subscriptions: simple, powerful and easy to get wrong Subscriptions are the most familiar loop, and for products used regularly they are usually the right starting point. They create predictable revenue, make planning easier and align the business with long-term customer success. But subscriptions have failure modes: Flat prices for variable costs. If some customers use far more than others — especially where AI costs scale with use — a flat price can make heavy users unprofitable. That is why MerchNivo combines a subscription with usage-based components. See Subscription plus usage: pricing logic for AI SaaS. Invisible value. A subscription for something that works quietly in the background risks being cancelled in a budget review. Products should make their value visible — through summaries, reports and reminders of what they did. Annual lock-in without annual value. Annual plans reduce churn on paper, but if customers feel trapped, the relationship suffers at renewal. Usage and credits: pricing that moves with value Usage-based pricing charges customers in proportion to what they use: per analysis, per report, per reading, per request. Credits are a friendlier version — customers buy a bundle in advance and spend it as they go. These models work well when: usage varies a lot between customers; the cost of serving each unit is significant (as with AI); customers want to try the product without committing to a subscription. ZodiVela combines subscriptions for regular users with credits and one-off premium readings for people who prefer to pay occasionally. The loop comes from giving people reasons to return — new readings, new formats — rather than from locking them in. Licences: recurring revenue on institutional calendars For education, the natural rhythm is the school year. EduRelia sells annual licences to schools in tiers — 100, 500 or 1,000+ students — alongside subscriptions for families. The school licence creates a predictable annual cycle, and renewal depends on whether teachers, students and families found real value during the year. Licence loops have their own discipline: onboarding at the start of term matters enormously, mid-year support determines renewal, and the end-of-year report is effectively the renewal pitch. We describe the model in B2B2C in education: family subscriptions and school licences. Repeat purchases: the commerce loop A direct-to-consumer brand like Noveniq does not have subscriptions in the usual sense. Its loop is repeat purchase: customers who bought once come back when they change devices, need a replacement or want something new. The ingredients still apply. The recurring problem is real (devices change regularly). The recurring moment of value is a new product that fits a new need. The low-friction path is a store that remembers nothing it should not, but makes reordering easy, communicates only when there is something genuinely useful to say, and treats customers well enough that they choose to return. See Running a direct-to-consumer brand as an operating business. Marketplace loops: two sides, two reasons to return A marketplace like CastLyra needs a loop on each side. For businesses, the recurring problem is that content production never stops: new campaigns, new products, new seasons. A subscription makes sense if the marketplace reliably delivers relevant, verified talent whenever they need it. For talent, the recurring problem is finding the next opportunity. A premium plan makes sense if it genuinely improves visibility or helps them manage requests better. Over time, completed bookings can carry a commission, aligning the marketplace's revenue with successful matches. The loops reinforce each other: more active businesses attract better talent, and better talent keeps businesses subscribed. The hard part is starting the loop at all — see Two-sided marketplaces: thinking about the cold-start problem and Marketplace monetisation. Report and partner loops: when the problem is episodic Some problems are not weekly or monthly. Deciding whether to study, work or live abroad is a significant, episodic decision. KeşifAtlası's core product — an eligibility test followed by a paid personal report — does not naturally repeat every month. The loop here comes from three places: Life stages. A family may explore options for study now, work later, and residence after that. Each stage is a new decision. Rule changes. When relevant rules change, an updated assessment has real value. Partners. People who decide to proceed often need further help — the report can connect them to relevant relocation partners, which creates referral revenue without the product itself needing to become a service business. Episodic products need to be especially careful about trust. A loop built on unnecessary upsells would undermine the core value. See From free test to paid report: designing a consumer funnel. Designing the loop before the product Why ask this question before building? Because the loop shapes the product in concrete ways. It decides what the product must remember. A subscription product needs a reason to show value repeatedly, which may require keeping history — carefully, within the product's data boundary. A report product needs to know when a rule relevant to a past report has changed. It decides what the product must measure. For a subscription, retention and engagement over time matter most. For credits, purchase frequency. For licences, usage across a term. The measurement plan follows the loop. It decides the pricing unit. Per store, per student, per report, per reading, per booking. Choosing the unit early avoids painful migrations later. It decides where AI costs must be controlled. If the loop depends on frequent AI-generated value, AI cost per customer must be sustainable at that frequency. See Cost discipline for AI products. Healthy loops versus manipulative ones There is a dark side to recurring revenue: products that are hard to cancel, subscriptions that renew silently at higher prices, "free trials" designed to be forgotten. These tactics can inflate revenue for a while. They also destroy trust, create complaints and invite regulatory attention. Our rules are simple: cancelling should be as easy as signing up; renewal reminders go out before annual renewals; price changes are announced in advance and explained; credits and prepaid balances are clearly shown; customers can export their data when they leave. A loop that works only because people cannot find the exit is not a loop worth building. Measuring whether a loop is working We track a small set of numbers for each company's loop: Retention over time for subscriptions and licences. Repeat rate and time between purchases for commerce and credits. Expansion: do customers grow into higher tiers or higher usage? Referral: do existing customers bring new ones? Cost to serve per customer, compared with revenue per customer. These numbers feed directly into the continue, stop or scale checkpoints. A company with strong first purchases but weak repeat behaviour has an acquisition engine, not a business — yet. Questions to ask about your own loop If you are designing a new product, try answering these honestly: How often does your customer have this problem? What will they notice from your product each week or month? What would make them stop paying — and how would you know before they do? Does your price move with the value you deliver and the cost you incur? If acquisition stopped for three months, what would revenue look like? The last question is the clearest test. A business with a real recurring loop survives a quiet quarter. A treadmill business does not. When the loop leaks: reading churn properly Every recurring business loses some customers. What matters is understanding why, and whether the reasons are fixable. We separate churn into a few types, because each calls for a different response. Churn after the first period. Customers leave soon after their first month, term or purchase. This usually means the promise on the website and the experience in the product did not match — the offer attracted people the product does not serve well, or onboarding failed to deliver the first moment of value. The fix is in messaging and activation, not in retention campaigns. Churn after a period of real use. Customers who used the product properly for a while decide to leave. This is the most informative churn, because these people understood the product. Talking to them — briefly, respectfully, without trying to talk them out of it — often reveals the next most important improvement. Churn caused by price. Sometimes the product is valued but the price no longer fits: a store's sales drop seasonally, a family's circumstances change. Flexible tiers, pausing options or usage-based pricing can keep these relationships alive without discounting the product for everyone. Churn that is healthy. Some customers leave because the problem is solved. A family that has completed a move no longer needs relocation reports. A student who has finished a course moves on. That is success, and the right response is to make it easy to come back — or to recommend the product to someone else — later. For each company we keep a short, regularly updated note on why customers leave, in their own words. It is one of the most useful documents a company can have, and it feeds directly into product priorities and the group's operating cadence. Subscription or usage: the trade-off in one table Most loop decisions end up somewhere between a flat subscription and pure usage pricing. Neither is better in general; each gives something up. Consideration Flat subscription Usage or credits Revenue predictability High; easy to forecast Lower; moves with customer activity Alignment with cost to serve Weak when usage varies widely Strong; heavy users pay more Customer's sense of control Lower; "am I using enough to justify this?" Higher; they pay for what they use Friction at the moment of use None Some; people think before spending Admin and support Simple Balances, top-ups and questions about charges The friction row is the one people forget. Usage pricing can quietly discourage the very behaviour the loop depends on — if every reading or analysis feels like spending money, customers use the product less and notice its value less. That is why several of our companies combine the two: a subscription that covers normal use comfortably, with usage or credits for the expensive edges. Mistakes we see in loop design Some mistakes come up often enough that we check for them in every new company's plan. Confusing notifications with value. A daily email is not a recurring moment of value unless the customer would miss it. Sending more messages to "drive engagement" usually drives unsubscribes. MerchNivo's briefing earns its place only if it contains something the merchant would otherwise have had to find themselves. Building the return reason on anxiety. In consumer products especially, it is tempting to bring people back through worry — about the future, about missing out, about a deadline that is not real. We rule this out. ZodiVela gives people reasons to return that are about reflection and enjoyment, never fear. See Entertainment positioning for consumer AI. Choosing the pricing unit for convenience. Pricing per seat is easy to bill, but if value grows with orders, students or requests, the unit fights the product. Migrating pricing units after launch is painful for everyone. Discounting to renew. A large renewal discount can save a customer this month and teach every customer to threaten cancellation. Better to fix the reason they wanted to leave, or offer a smaller plan that fits. A loop that depends on one channel. If repeat revenue depends entirely on a single marketplace, platform or partner, the loop is rented. Owned relationships — email with proper consent, accounts, a good product — keep it yours. A one-page loop worksheet Before any build starts, the loop is written into the company's source-of-truth document as a short worksheet. It has seven fields: The recurring problem, and how often the customer has it. The recurring moment of value, described as the customer would experience it. The pricing unit, and why it tracks value. The cost to serve one unit, including AI, at expected usage. The renewal or repeat moment — when and how the customer decides to pay again. The exit: how a customer cancels, exports data or stops, in a sentence. The measure: the one or two numbers that show the loop is working. Filling it in takes an hour. Leaving a field blank is informative in itself — it usually means that part of the loop has not been thought through yet. For EduRelia, for example, the renewal moment is not a monthly charge but a school's budget decision near the end of the academic year, and the worksheet makes clear that the evidence for that decision must be collected all year, not in the final month. The worksheet is revisited at each checkpoint. Loops change as companies learn — a report product may discover that partners matter more than expected, or a subscription product that a credit option brings in customers who would never subscribe. Writing the change down keeps the whole team working to the same model, and makes it easier to see which products in the group share the same shape of problem. Summary Recurring revenue is not a pricing trick; it is a consequence of solving a recurring problem and delivering value the customer notices, repeatedly and fairly. At Oryvelon we design the loop before the product — subscriptions where use is regular, usage and credits where it varies, licences on institutional calendars, repeat purchases in commerce, and partner loops where the problem is episodic. The result is a group of companies whose revenue compounds rather than resets. You can see the full list on the Companies page. ## Capital allocation inside a company builder URL: https://oryvelon.com/insights/capital-allocation-in-a-company-builder/ Category: Company Building In a single-company startup, capital allocation is mostly a question of how much to spend and on what. In a company builder there is a harder question on top: which company? Every Oryvelon company competes for the same limited resources. A month of engineering time spent improving ZodiVela is a month not spent on MerchNivo. An afternoon the founders spend on a CastLyra partnership is an afternoon not spent on EduRelia schools. Getting these decisions right is, arguably, the most important thing a company builder does. This note explains how we approach it. Three kinds of capital When people hear "capital allocation" they think of money. In a company builder, money is only one of three resources, and often not the scarcest. Money pays for acquisition, infrastructure, AI usage, tools, freelancers and sometimes hires. Build capacity — engineering, design and content time — determines how fast products improve. Founder attention is the scarcest of all. Partnerships, key customer relationships, hiring, strategy, brand decisions and the hardest problems all need it, and there is only so much of it in a week. We allocate all three deliberately. A company might receive a modest budget but a lot of founder attention during a critical partnership period, or the reverse once it runs smoothly. Principle one: evidence over enthusiasm Every founder has favourite ideas. The danger in a company builder is that favourites receive resources regardless of results, while less exciting companies with better numbers are starved. Our protection against this is the continue, stop or scale process. Each company has KPIs written before launch and checkpoints at 30, 60 and 90 days. Allocation decisions are based on what those checkpoints show: Are the right customers arriving and activating? Are they coming back? Are they paying, and staying? Does each customer generate more than they cost to acquire and serve? Is the team learning faster than the problems grow? Enthusiasm still matters — it keeps teams going through hard months — but it does not decide budgets. Principle two: visible per-company economics You cannot compare companies fairly if you cannot see their economics separately. A group invoice that bundles hosting, AI usage, tools and payroll across companies makes it impossible to say which company is healthy. At Oryvelon, each company's costs are attributed to that company as far as practical: hosting and infrastructure; database and storage; AI usage, through separate AI projects and budgets; email and messaging; tools and subscriptions; a fair share of shared costs, allocated transparently. Next to those costs sit revenue, gross margin and the key customer metrics. That view is what makes allocation decisions honest. See Unit economics per product in a shared-infrastructure group. Principle three: staged commitments We rarely make large, irreversible commitments early. Instead, investment grows in stages as evidence accumulates. Stage Typical commitment Evidence required to move on Validation Founder time, a phase-zero site, small experiments Real problem, willingness to pay, a reachable audience First version Focused build capacity on the shared foundation Activation and early retention Early growth Modest acquisition budget, improvement work Retention holds; unit economics positive or clearly improving Scale Larger budgets, more build capacity, partners Retention, margin and quality hold as volume grows Each stage is a smaller bet than it would be for a standalone startup, because the foundation already exists. That lets us try more ideas at the early stages while being stricter about which ones advance. Principle four: move resources away, not just towards Allocating to winners is the easy half. The hard half is taking resources away from companies that are not working — or that are working but have reached a plateau where more investment will not change much. We make that easier in three ways: Pre-agreed checkpoints mean reductions are expected, not personal. Clear definitions of "steady state" let a company that is healthy but not growing fast run efficiently on a smaller share of attention. Responsible stopping preserves value when a company or initiative ends — domains, content, code and learning stay in the group. See Continue, stop or scale. A portfolio where every company always gets a little more is not being managed. It is being hoped for. Principle five: protect the shared foundation There is one allocation that is easy to neglect because it does not belong to any single company: the shared operating core. Domain and email security, deployment standards, the AI gateway, measurement, access reviews and cost tracking all need ongoing investment. If the core is starved, every company slowly inherits problems: outdated standards, security gaps, tools that no longer fit. So a fixed, modest share of build capacity goes to the core every cycle, and improvements are prioritised by how many companies they help. See What is a digital company builder?. How a typical allocation review works We review allocation on a regular rhythm, aligned with the group's operating cadence. The review is short and follows the same structure every time. Each company's snapshot. Key metrics against targets, economics, recent checkpoint decisions, and one paragraph from the company's owner on what they need most. The core's snapshot. Open risks, planned improvements, and which companies would benefit. Proposed changes. More, the same or less money, build capacity and founder attention for each company, with reasoning. Decision and record. Changes are agreed and written down, with the date of the next review. The written record is important. Months later, when a decision looks obviously right or wrong, we can see what we knew at the time and learn from it. How it plays out across very different companies The companies in the group have different natural rhythms, and allocation respects that. Commerce and services businesses such as Noveniq and WeAreMedia have established models. Allocation focuses on efficiency and steady improvement rather than exploration. Software and AI products such as MerchNivo and KeşifAtlası need concentrated build capacity in early stages and careful attention to AI unit economics as they grow. Marketplaces such as CastLyra often need founder attention more than money early on — relationships on both sides matter enormously before the network effect starts. Education in EduRelia follows the school calendar; allocation peaks before and at the start of terms. Consumer products such as ZodiVela benefit from rapid experimentation budgets once retention is proven. Service-and-education brands such as Sinem Keser Beauty Academy grow through reputation, content and repeat relationships, where brand investment and quality of delivery matter most. Mistakes we try to avoid A few allocation mistakes are common enough that we name them explicitly: Equal shares. Dividing budget evenly feels fair and avoids conflict. It also guarantees under-investment in the best opportunities. Newest-idea bias. New ideas are exciting; established companies are less so. Evidence should decide, not novelty. Scaling before retention. Acquisition spend on a product that does not retain users multiplies losses. Ignoring attention. A company with enough money but no founder attention can stall just as badly as one without money. Starving the core. Short-term savings on shared foundations become long-term costs everywhere. Decisions without records. Without written reasoning, the group cannot learn from its own allocation history. What this means for partners and investors If you are considering partnering with or investing in an Oryvelon company, the allocation discipline has two practical implications. First, the numbers exist. Because each company's economics are tracked separately, there is a clear picture of revenue, costs and customer behaviour for that company alone. Second, group support is deliberate. When a company receives more of the group's resources, it is because the evidence justified it — and the conditions for continuing that support are written down. An illustrative allocation round To make the process concrete, here is a simplified, illustrative round — the kind of reasoning that happens in a review, not a report of specific numbers. Imagine three companies reaching checkpoints in the same month. Company A has modest growth, but almost every customer who activated two months ago is still active, and several are paying for higher usage tiers. Its unit economics are positive. The owner asks for more acquisition budget and a small amount of engineering time for the features retained customers request most. The group agrees to a staged increase — a larger budget for one cycle, with a review date and a guardrail: retention must hold at current levels. Company B had an excellent launch month but weak return behaviour. The owner proposes a bigger marketing push to "keep momentum". The group declines additional acquisition spend and instead moves engineering time to onboarding and the core experience, with a clear question for the next checkpoint: do the people who arrive now come back? Company C is stable, profitable and growing slowly. It does not need more money; it needs less founder attention so that attention can go elsewhere. The group documents how it runs, sets a steady-state budget and moves its review to a lighter rhythm. Nobody "won" or "lost" in that round. Each company received what the evidence suggested it could use well. Signals that trigger an off-cycle review Most allocation happens on the regular rhythm, but some events justify an immediate review: a sudden change in retention, positive or negative; AI or infrastructure costs rising faster than usage; a partnership or distribution opportunity that needs a quick answer; a security or data incident that requires resources to be redirected; a regulatory or platform change that affects a company's model; a company hitting a hard budget limit. Off-cycle reviews are short and focused on one decision. They follow the same principle: evidence first, then a written decision. Scheduling founder attention Because founder attention is the scarcest resource, we treat it like a budget too. Each cycle, the founders decide roughly how their time will be split across companies and the core, based on where their involvement makes the biggest difference. A few patterns help: Concentrate rather than spread. A company that needs founder attention usually needs a lot of it for a short period — a launch, a key partnership, a pricing change — rather than a little of it forever. Delegate steady state. Once a company runs smoothly, its day-to-day decisions should not wait for founders. Protect deep work on the core. Standards, security and architecture decisions need uninterrupted time; they get scheduled, not squeezed in. Say no explicitly. When attention goes to one company, it is worth being clear about what is not getting attention this cycle. Allocation and risk Allocation is also a risk decision. Concentrating resources in one company increases the group's dependence on it. Spreading them thinly reduces the chance that any company succeeds. We try to keep a balance: established businesses with steady revenue provide stability, while newer products receive the focused investment they need to prove themselves. Because each company is designed to stand alone, a problem in one company does not threaten the others technically — the shared layer is infrastructure and standards, not a single platform that could fail for everyone. Ten questions every allocation review answers To keep reviews consistent, each one ends by answering the same ten questions in writing. Which company has the strongest evidence of a recurring, paying customer base? Which company is closest to a breakthrough if it receives more build capacity? Which company needs founder attention most urgently, and for how long? Which company can move to a lighter, steady-state rhythm? Where are costs growing faster than value, and what is the fix? Is any company being kept alive by hope rather than evidence? What does the shared core need this cycle, and which companies benefit? Which risks — security, regulatory, platform, concentration — need resources now? What did the previous round get wrong, and what do we change because of it? When is the next review, and what do we expect to see by then? Answering all ten takes discipline, especially the sixth and ninth. They are the questions that make a company builder more than a collection of projects. Worked example: when AI costs jump at one company Here is an illustrative off-cycle review of the kind described above — a pattern, not a report. Imagine a relevant visa rule changes in a destination many KeşifAtlası users are exploring. Interest rises, more people take the free eligibility test, and more existing report holders come back to check how the change affects them. AI usage climbs sharply in a few days, and the company approaches its monthly AI budget limit. The easy responses are both wrong. Raising the budget without thought rewards a cost problem with more money. Switching off AI explanations hurts users exactly when the product is most useful. Instead, the review asks three questions: Is this demand we want? Yes — it is the product doing its job at a moment of real need. Where is the cost going? Often the same explanation of the same rule is being generated many times. Eligibility itself comes from the verified rules engine; the model only explains the result, so identical explanations can be cached safely. See Latency and caching for AI. Is the most capable model needed for every step? Simple rephrasing and classification can go to smaller, cheaper models, with the capable model kept for the harder explanations. See Model routing. The decision might be a temporary budget increase with a fixed end date, alongside engineering time for caching and routing. The written record says what the extra money is for and when it stops. Allocation here is not a yes or no to spending; it is a choice about what kind of spending fixes the problem. Allocating shared costs fairly Shared costs are where per-company economics usually get distorted. If the group simply divides the core's costs equally, a small, early company looks unprofitable and a large one looks cheaper than it is. Both views lead to bad allocation. We use a simple order of preference: Method When we use it Example Direct attribution The cost can be tagged to one company A company's own database, AI project, sending domain Measured usage The cost is shared but usage is measurable Requests routed through the AI gateway, per company Agreed proportion Neither of the above is practical A shared tool or a slice of core engineering time Whatever the method, it is written down and applied the same way every cycle. Changing the allocation formula to make a favoured company look better is exactly the kind of drift the process exists to prevent. When the formula itself seems unfair, we change it openly, for everyone, and note the date. Cheap bets that pay twice Some allocations produce value in more than one place, and we look for them deliberately. Noveniq is the clearest case. It is a real store with its own customers and economics, and it has to stand on those. But running it also teaches the group what store operators face every day — the questions that make MerchNivo more useful. Time spent improving Noveniq's operations is judged first on Noveniq's own results; the learning for MerchNivo is a secondary return, not an excuse for weak numbers. See A real store as a testbed. The same logic applies to shared standards. A privacy or security improvement built for EduRelia, where the bar for child data is highest, often becomes the default for other companies. Investment made where requirements are strictest tends to travel well. A pre-review checklist for company owners Before each allocation review, the owner of each company prepares: [ ] Current KPIs against the targets written at the last decision. [ ] Costs for the period, broken down, with anything unusual explained. [ ] One specific request — money, build capacity or founder attention — with what it would change. [ ] What the company would do if the answer were no. [ ] One thing the company should stop doing to free resources. The fourth and fifth items matter most. An owner who has thought about the "no" case and found something to stop is making allocation easier for everyone, including their own company. Summary Capital allocation is where a company builder earns or loses its value. At Oryvelon we treat money, build capacity and founder attention as resources to be allocated deliberately; we let evidence rather than enthusiasm decide; we keep each company's economics visible; we stage commitments; we are willing to move resources away; and we protect the shared foundation that every company relies on. It is not glamorous work, but it determines which companies thrive. ## Designing every company so it could stand alone URL: https://oryvelon.com/insights/designing-companies-to-stand-alone/ Category: Company Building Most companies in a group never leave it. But the ones that could are healthier than the ones that could not. At Oryvelon we design every company so that it could stand alone: raise its own capital, bring in a strategic partner, be spun out, or be sold. We have no plans to do any of those things today. We design for them anyway, because separability forces a set of good decisions that make each company better right now. This note explains what "could stand alone" means in practice, and the choices we make from day one to make it true. Why separability matters even if you never separate There are three reasons to care about separability before it is needed. It keeps options open. Opportunities arrive unexpectedly: a partner who wants a stake in one company, an investor interested in one market, a buyer for a business that fits their portfolio. A company that can be separated in weeks can take those opportunities. One that needs a year of untangling usually cannot. It protects customers. The same boundaries that make a company separable also keep its customers' data where it belongs. If a company's users live in its own database, they are not exposed to the others — and they can move with the company if it ever leaves. See Shared infrastructure, separate data. It makes each company easier to understand. When a company owns its code, data, infrastructure and economics, anyone — a new team member, an auditor, the founders themselves — can see exactly what it is and how it performs. What makes separation hard It helps to know what goes wrong in groups that did not plan for separation. The common problems are: Shared user databases. Customers of several products live in one set of tables. Separating them means migrating data, merging or splitting accounts, and hoping nothing breaks. A shared identity system. One login for everything. Removing a product means re-issuing credentials to its users and rebuilding authentication. Code with product-specific exceptions. Shared modules full of "if product == X" branches. Extracting one product means understanding and rewriting large parts of the shared code. Mixed vendor accounts. One analytics property, one email account, one payment account, one AI key for everything. Separating them means rebuilding history and integrations. Blended finances. Costs and revenue that were never attributed per product. Nobody can say what the company earns or costs on its own. Undocumented decisions. Nobody remembers why things are the way they are. Each of these is a reasonable shortcut on the day it is taken. Together they make separation a painful, expensive project. We avoid all of them. What each Oryvelon company owns From its first day, each company in the group owns a complete set of its own assets. Asset Owned by each company Domains Registered for the company, locked, auto-renewing Website Its own site on its own domain Code Its own repositories and release cycle Database Its own database containing only its data Users Its own accounts; no shared identity layer AI Its own AI project, keys, prompts, knowledge and budget Email Its own sending domain with SPF, DKIM and DMARC Analytics Its own properties and measurement Payments Its own payment setup where it charges customers Legal pages Its own privacy notice and terms Economics Its own attributed costs and revenue The group provides standards, tooling and people who know how to use them. It does not own a platform the companies depend on. Shared code that does not trap anyone Companies still benefit from shared code — it would be wasteful for each to write its own form components, logging helpers or security headers. The trick is to share code in a way that never becomes a trap. Brand-neutral. Shared packages contain no product logic, no brand styling and no business rules. They are building blocks that each company assembles in its own way. Versioned. Each company depends on a specific version of shared code and upgrades when it is ready. A change in a shared package cannot force every company to release at once. Replaceable. Because shared packages are small and focused, a separated company could keep using them, fork them or replace them without rewriting its product. We discuss the boundary between shared and product code in Why we don't merge our products into one app. Infrastructure that can be handed over Infrastructure is organised so that each company's resources are grouped together and could be transferred as a unit: hosting projects, database projects, storage, background jobs, monitoring and AI projects. Naming standards — part of step six in our new-company protocol — make it obvious which resources belong to which company. Access to each company's resources is granted per company and per role. That makes it straightforward, in a separation, to hand over administrative control to a new owner and remove the group's access cleanly. See Least-privilege access for small teams and contractors. Records that make due diligence easy When a partner or buyer looks at a company, they ask predictable questions. We keep the answers ready as a matter of routine, not as a special project. What is this company? The source-of-truth document. See Why every company starts with a source-of-truth document. How does it make money, and what does it cost? Per-company economics, tracked from day one. See Unit economics per product. What data does it hold? The data sensitivity classification and data inventory. Who has access to what? Access lists, reviewed quarterly. What does it depend on? A list of vendors, tools and shared packages, with versions. How is it secured? Domain, DNS, email authentication, security headers, secrets management and incident history. What has it learned? Checkpoint reviews and the decisions made at each one. A company with these records can go through due diligence in a fraction of the usual time. Customer relationships that belong to the company Separability is not only technical. Each company's relationship with its customers belongs to that company: customers sign up with the company, under its terms and privacy notice; they receive communication from the company's domain and in its voice; their support conversations happen with the company; their consent — for product use and for marketing — is recorded by the company. If the company ever left the group, those relationships would move with it intact. Nobody would discover that their data had been sitting in someone else's system. Where the group relationship lives A company that stands alone still benefits from being part of Oryvelon, and we are happy for that to be visible. The relationship appears in natural places: a short "An Oryvelon company" note in the footer or on the about page; a profile on oryvelon.com's Companies page; mentions in press or partnership contexts where the group is relevant. What we avoid is making the group part of the company's core identity — the Oryvelon mark is not part of any company's logo, and the group name is not placed larger than the company's own brand. A company's identity should survive intact if the footer line ever changes. The cost of designing this way Separability has costs. Some infrastructure is duplicated. Customers who use two products have two accounts. Each company has its own analytics to maintain. Shared code must be kept brand-neutral and versioned, which takes discipline. We think these costs are small compared with the alternative. And many of them are costs a well-run company should bear anyway: its own data boundary, its own measurement, its own clear economics. A separability checklist For any company in the group, we can ask: [ ] Could we list every resource that belongs to this company in an hour? [ ] Does its database contain only its data? [ ] Could its users keep their accounts if it left the group? [ ] Are its AI keys, prompts and budgets its own? [ ] Does it send email from its own authenticated domain? [ ] Are its costs and revenue visible on their own? [ ] Is its shared code versioned and replaceable? [ ] Are its access lists current? [ ] Is its source of truth up to date? If every box is ticked, the company could stand alone. That is the standard every Oryvelon company is held to. Three separation scenarios It helps to think through what "standing alone" would mean in the situations where it could actually happen. A strategic partner takes a stake. A partner — for example, a business that distributes to the same customers — invests in one company. Nothing needs to move technically. What matters is that the company's economics, data and governance are clear enough for the partner to understand exactly what they are investing in, and that the group's shared services to the company can be described and priced fairly. The company raises its own capital. Investors want to see a company, not a slice of a group. The records described above — source of truth, economics, data inventory, access lists, vendor list — are exactly what they ask for. The company may continue to use group standards and shared packages under a clear agreement, or gradually take them in-house. The company is sold. This is the fullest test. Domains, repositories, infrastructure projects, AI projects, analytics properties, email configuration, payment accounts and vendor contracts transfer to the new owner. Access for group staff is removed. Customers continue with the same company under the same terms. Because none of the company's data or identity lives in shared systems, there is nothing to extract — only things to hand over. In all three scenarios, the work is administrative rather than surgical. That is the goal. What would stay with the group Separation also clarifies what belongs to the group rather than to any single company: the shared standards and documentation; the operating core's tools and know-how; the group brand and oryvelon.com; lessons learned, written into standards so other companies keep benefiting. A company leaving would take its assets with it. The group would keep the knowledge of how to build the next one. How separability shapes everyday engineering Designing for separation is not a special project. It shows up in small, everyday decisions: A developer adding a feature that needs data from another company is told no — and the feature is redesigned to work within the company's own boundary. A new tool is set up per company rather than as one account for the group. A shared package that starts to accumulate product-specific logic is split, and the product logic moves back into the product. A new environment variable or secret is created in the company's own projects, never in a group-wide store that several companies read. Costs from a new vendor are tagged to the company that uses them from the first invoice. None of these decisions is dramatic. Together they are what keeps each company whole. You can see the companies we hold to this standard on the Companies page, and the operating model behind them on the Build page. Questions partners ask, answered briefly "If we work with one of your companies, are we also working with the others?" No. Agreements, data and relationships are with the specific company. "Does the company depend on a group platform that could change under us?" No. It depends on versioned, brand-neutral packages and documented standards. Its product, data and infrastructure are its own. "Can we see this company's numbers on their own?" Yes. Costs and revenue are attributed per company from the start. "Who owns the domain and customer relationship?" The company does, with the group ensuring the domain is secured and renewed. "What happens to customer data if the company changes ownership?" It stays with the company, under the same privacy commitments, and moves with it. It was never stored anywhere else. These answers are short because the design makes them simple. That simplicity is the practical value of building every company to stand alone. Worked example: what a hand-over would involve To test whether separability is real, it helps to walk through an actual sequence. Take KeşifAtlası as a hypothetical case: imagine a partner wanted to take full ownership. Nothing like this is planned; the point is to check that every step is administrative. Inventory. List every resource tagged to the company — the naming standards make this a query, not a search. Access freeze. Pause non-essential changes, and record who currently has access to each system. Domains and DNS. Transfer the registrar account for kesifatlasi.com, then confirm DNS records and email authentication (SPF, DKIM, DMARC) still pass after the move. See Domain, DNS and email security. Code and infrastructure. Transfer repositories, hosting and database projects, storage and background jobs as whole units. AI project. Hand over the company's AI project, prompt versions and knowledge. The verified rules database goes with it, because it is the company's core asset. Secrets. Rotate every key and credential after transfer, so nothing the group once held still works. See Secrets management basics. Analytics, payments and vendors. Move properties, payment accounts and contracts, each already registered to the company. Remove group access. Close group staff access last, once the new owner confirms control. Tell customers what changes — ideally very little: same product, same domain, same privacy commitments. The shared AI gateway is the one piece that does not move. The company either keeps using it under an agreement or points its calls at its own setup. Because the gateway never merged its prompts, keys or users with anyone else's, switching is a configuration change. Mistakes that quietly erode separability Separability rarely fails in one big decision. It erodes through small, convenient ones. Domains in a personal account. A domain registered by a founder or contractor under their own login is a separation problem waiting to happen. Every domain sits in a company-controlled account. One marketing list for every brand. Collecting emails into a group-wide list feels efficient, but consent was given to a specific company. A shared list is both a separation problem and a consent problem. See Marketing consent vs product consent. A "temporary" data join. An analysis that copies two companies' customer data into one place "just this once" tends to become permanent. Shared test data. Test environments populated with another company's real data blur boundaries in the places nobody reviews. See Environments and test data. One payment account for convenience. Revenue that lands in a shared account has to be untangled later, along with refunds and disputes. Tenancy inside a shared app. Keeping companies apart by a column in one database is weaker than separate databases. We explain the difference in Tenant isolation explained. Each one saves an hour today and costs weeks later. Naming them makes them easy to refuse. Testing separability without separating We do not wait for a real separation to find out whether a company could stand alone. Once a year, each company goes through a short drill: rebuild the product from its own repositories into a clean environment, using only its own configuration; produce the full resource inventory and compare it with what is actually running; confirm that no credential used by the company is stored outside its own projects; check that its source of truth, access list and vendor list match reality. Anything the drill finds is fixed as ordinary work. A drill that finds nothing is the goal, and it becomes more likely every year the discipline holds. Summary Designing companies to stand alone is a set of ordinary, repeated choices: each company owns its domains, code, database, users, AI project, email, analytics, payments and economics; shared code stays brand-neutral, versioned and replaceable; records are kept ready for due diligence; and small shortcuts that blur boundaries are refused before they become habits. We have no plans to separate any company from Oryvelon, but building as if we might keeps customer data where it belongs, makes every company easier to understand and run, and turns any future separation into a hand-over rather than a surgery. ## Why every company starts with a source-of-truth document URL: https://oryvelon.com/insights/source-of-truth-documents/ Category: Company Building Every company eventually argues with itself. A designer proposes a feature that contradicts the pricing model. A marketer writes copy that describes the product as something it is not. An engineer adds a data field that the privacy plan never allowed. Nobody is wrong on purpose. They are working from different memories of what the company was meant to be. At Oryvelon, every company starts with a source-of-truth document — step five of our new-company protocol. It is short, it is written before code or design, and every later document is expected to agree with it. It is one of the simplest practices we have and one of the most valuable. What "source of truth" means In software, a source of truth is the one place where a piece of data is authoritative. Other systems may copy it, but when they disagree, the source of truth wins. A source-of-truth document applies the same idea to a company's intent. Product specs, marketing briefs, technical plans, sales decks and website copy may all describe the company. When they disagree with each other, the source of truth decides. When they disagree with the source of truth, either they are wrong or the source of truth needs a deliberate update. The group itself works this way. Oryvelon has a master project document that defines the group's role, its companies and domains, the boundaries between them, and the technical and operating standards they share. Company-level documents must not contradict it. What goes into it A company's source of truth is usually a few pages. It covers the same headings every time, which makes companies easy to compare and documents easy to navigate. Identity. The company's name, domain and one-sentence description. Customer and problem. Who the company serves and what problem it solves, in the customer's language. Category. Software, marketplace, AI product, commerce, services, education — and what that implies. Revenue mechanism. Who pays, for what, and how revenue repeats. See Designing a recurring revenue loop. What the company will not do. Often the most useful section. Data classification and rules. What data it holds, how sensitive it is, and the resulting rules for isolation, access, retention and AI use. AI principles. Which facts must come from verified systems, and what AI is allowed to do. See AI explains, verified data decides. Relationship to the group. What the company shares with Oryvelon, what it does not, and how the relationship is shown publicly. Brand voice and positioning. How the company speaks and what claims it must never make. Success measures. The KPIs and checkpoint dates. See Continue, stop or scale. The "will not do" section If we had to keep only one section, it might be this one. Positive descriptions are easy to stretch; explicit boundaries are not. Some examples of the kind of statements that belong here: Oryvelon will not sell agency services directly; those belong to WeAreMedia. ZodiVela will not present readings as predictions or as medical, legal or financial advice. KeşifAtlası will not let an AI model decide eligibility; the verified rules database does. EduRelia will not use student data in any consumer product or for marketing. MerchNivo will not use one store's data for another store or another product. CastLyra will not share talent contact details outside its access rules. Each of these lines settles a whole category of future debates in advance. When someone proposes a clever idea that crosses one, the conversation is short. Why it saves time A source-of-truth document can look like overhead. In practice it pays for itself quickly. It resolves disagreements fast. When two people disagree about direction, the first question is "what does the source of truth say?" Often that settles it. If it does not, the disagreement becomes a clear proposal to change the document, which is a better conversation than an argument about opinions. It onboards people faster. A new team member, freelancer or partner can read a few pages and understand what the company is, what it is not and why. That is far quicker than piecing it together from meetings and chat history. It keeps AI tools aligned. When we use AI assistants to help draft content, specs or code, the source of truth is part of the context they are given. That keeps generated work consistent with the company's intent — and makes it obvious when something does not fit. It prevents slow drift. Companies rarely change direction in one dramatic decision. They drift, one reasonable-sounding exception at a time. A written reference makes drift visible. It makes the company separable. Anyone evaluating a company — a partner, investor or acquirer — wants to know what it is and what it has committed to. The source of truth answers that directly. See Designing every company so it could stand alone. Keeping it alive A document that is written once and never touched becomes a museum piece. We keep sources of truth alive with a few habits. Short beats complete. A few pages that everyone reads are worth more than fifty that nobody does. Details belong in specs; the source of truth holds decisions. Versioned, with a change log. Every change records what changed, when, and why. Months later, the history tells the story of how the company evolved. Updated deliberately. Changing the source of truth is a decision, usually made at a checkpoint or allocation review. It is never edited casually to make a proposal fit. Referenced constantly. Specs, briefs and plans link to the relevant section. If a document cannot point to the part of the source of truth it serves, that is a warning sign. Reviewed at checkpoints. At each continue, stop or scale checkpoint we ask whether the source of truth still describes the company accurately. If the company has learned something fundamental, the document changes. A hierarchy of documents Sources of truth sit at the top of a simple hierarchy: The group master document — Oryvelon's role, companies, boundaries and standards. Each company's source of truth — identity, customer, revenue, boundaries, data rules. Specifications and plans — product specs, technical designs, marketing plans. Working documents — tickets, drafts, meeting notes. Each level must agree with the levels above it. When a lower-level document needs something the higher level forbids, the higher level is either updated deliberately or the lower-level document changes. Common objections "We're too small for this." Small teams benefit most, because they have the least time to repeat the same arguments. "It will be out of date immediately." Only if nobody updates it. Keeping it short and reviewing it at checkpoints solves most of this. "It kills creativity." It channels it. Knowing what a company will not do makes it easier to be bold within the space it will. How to write one If you want to try this for your own company, start with a blank page and the ten headings above. Give yourself an hour. Write a paragraph or less under each. Where you cannot, you have found an unresolved question that deserves attention. Share the draft with the people who work on the company, invite disagreement, and settle it in writing. Then use it. Link to it from every spec. Open it at the start of every planning session. Change it when you learn something fundamental, and write down why. What the group-level source of truth looks like Oryvelon's own master document is a good illustration of the format, because it covers everything the companies inherit. Its sections, in order, are roughly: The core decision and unchangeable rules. Oryvelon is a company builder and venture group — not an agency, not a single SaaS product, not a portfolio showcase. Oryvelon's mission and what it will not do. Build, own and operate companies; never run each company's daily customer operations centrally or merge products into one application. The exact company and domain list. Every brand, its category, its main job and its revenue model. Brand boundaries between companies. How each company relates to the others, including which authority sites may send qualified traffic where. Website information architecture. The pages oryvelon.com has, and what each is for. The design direction. Minimal, editorial, technology, premium, global — and what to avoid. The "An Oryvelon company" standard. Where the group relationship may appear, and where it must not. The shared technical core. Which services provide DNS, code hosting, deployment, databases, AI access, background jobs, email, analytics and monitoring. The AI gateway. Routing, prompt registry, budgets, safety, observability, fallbacks and caching — with the critical rule that AI is not the source of business truth. Data isolation and privacy boundaries. The table of what each company keeps separate. Domain, DNS and email standards. Repository and code organisation. Environments. Local, preview, staging and production, and which data each may contain. AI governance. Separate projects, budgets, server-side keys, versioned prompts, structured outputs, evaluation, fallbacks, human escalation. Shared and non-shared libraries. Analytics and UTM standards. Search, entity and AI visibility. Security, access and roles. Finance and cost control. The new-company protocol. Content policy for insights. Implementation order and QA criteria. That may sound long, but each section is only a few lines. It fits in a document you can read in one sitting — and every company-level source of truth starts by inheriting it. Source of truth, strategy deck, wiki: what is the difference? Companies accumulate many kinds of documents. It helps to be clear about what the source of truth is not. A strategy deck persuades. It is written for an audience — investors, partners, a board — and it emphasises the most compelling parts of the story. A source of truth describes; it includes the boundaries and the uncomfortable constraints. A wiki stores. It grows in every direction, contains contradictions and is rarely read from start to finish. A source of truth is short, curated and internally consistent. A product spec details. It explains how a specific feature will work. A source of truth explains why the product exists and which rules every feature must respect. A brand guide styles. It governs how the company looks and sounds. A source of truth sets the positioning and claims the brand guide must support. All of these documents are useful. The source of truth is the one they should all agree with. Signs a source of truth is failing We watch for a few symptoms that suggest a company's source of truth has stopped doing its job: people answer "what is this company?" differently from the document; specs and briefs rarely link to it; the last change was long ago, even though the company has clearly learned important things since; a decision was made that contradicts it, and nobody noticed; new team members are told "don't bother reading that, it's out of date". When we see these signs, the fix is usually quick: a short session to reconcile the document with reality, record the decisions that were made informally, and remove anything that no longer applies. Using the source of truth with AI assistants We use AI assistants to help draft specs, content, code and plans. The source of truth has become one of the most important inputs we give them. When an assistant receives a company's source of truth as context, its drafts start from the right assumptions: the right customer, the right tone, the right boundaries. It is far less likely to propose a feature that crosses a data rule or a claim the brand must never make. And when it does, the conflict is easy to spot, because the rule is written down and can be pointed to. The same applies to automated work. Scheduled jobs that update content or check a site's health are given the relevant source of truth and content rules as part of their instructions. That keeps automation aligned with human intent, even when nobody is watching in real time. A worked example: KeşifAtlası The ten headings are easier to understand with one company filled in. Here is an illustrative, shortened view of how the source of truth for KeşifAtlası reads under a few of them. The wording is simplified for this article, but the shape is real. Identity. KeşifAtlası helps people understand visa and relocation options, starting from Türkiye. It explains which routes may fit their situation, based on a verified rules database. Customer and problem. People who are considering a move or a long stay abroad and cannot tell which rules apply to them. Official information is scattered, written for lawyers and changes without notice. Revenue mechanism. A free eligibility test, followed by a paid, more detailed report for people who want it. The logic of that step is described in From free test to paid report. Will not do. Let a model decide eligibility. Promise an outcome that depends on an authority. Answer a question the rules database does not cover as if it did. Data rules. Collect only what the eligibility rules need. Keep answers separate from marketing lists. Delete test data on a defined schedule. See Data retention and deletion. AI principles. The rules engine produces the result; AI explains it in plain language and must reference the rule it is explaining. Hard or unusual cases go to a person, as described in Human escalation in AI products. Notice how much this settles. A proposal to "let the model estimate eligibility when a rule is missing" is answered before anyone builds it. A request to reuse test answers for a newsletter segment is answered too. Nobody needs a meeting. How a change actually happens Changing the source of truth should be possible, but never quiet. Our process is short. Someone writes a proposal. One paragraph: what should change, which section, and the evidence behind it. Evidence usually comes from a checkpoint, customer conversations or a problem that keeps recurring. The owner of the company and a founder review it. Most proposals are decided within days. Some are sent back for more evidence. The change is made with a log entry. Date, section, old wording, new wording, reason. Dependent documents are checked. If the revenue section changed, the pricing page, onboarding copy and analytics events are reviewed for anything that now disagrees. Step four is the one teams skip. A source of truth that changes while the specs below it stay the same creates exactly the confusion it was meant to prevent. A typical, hypothetical log entry looks like this: Field Entry Section Revenue mechanism Change Paid report now includes one follow-up question answered by a person Reason Buyers repeatedly asked a single clarifying question after reading the report Documents to check Pricing page, report template, escalation queue rules Common mistakes when writing one We have made most of these ourselves at some point. Writing aspirations instead of decisions. "We want to be the most trusted platform in our category" tells nobody what to do. "We will not show a result without the rule that produced it" does. Leaving the data section vague. "We take privacy seriously" is not a rule. Name the data, its sensitivity, who can access it, how long it is kept and whether AI may see it. Copying the group document. A company's source of truth inherits the group's rules; it does not need to repeat them. It should record what is specific to this company. Letting it grow into a spec. When a section starts describing screens and buttons, the detail belongs in a product document that links back. Having no owner. Every source of truth has one named person responsible for keeping it accurate. Without that, everyone assumes someone else will update it. A quick checklist Before a new company moves into build, its source of truth should pass a short test: Can a newcomer describe the company in one sentence after reading it? Is it clear who pays, and for what? Does the "will not do" section contain at least a few specific, testable statements? Is every category of data named, with a rule for access, retention and AI use? Does it say which facts come from verified systems? Are the checkpoint KPIs and dates written down? Is there an owner and a change log? If any answer is no, the gap is cheaper to close now than after the first release. Summary A source-of-truth document is a small investment with a large return: faster decisions, faster onboarding, less drift, better alignment with AI tools and a company that is easier to understand from the outside. Every Oryvelon company has one, and the group has one too. It is the first thing we write, and the thing we return to most often. ## Brand architecture for a multi-company group: parent brand vs. product brands URL: https://oryvelon.com/insights/naming-domains-and-brand-architecture/ Category: Company Building Every group that runs more than one business has to decide how its brands relate to each other. Should every product carry the parent's name? Should the parent disappear entirely? Or something in between? The choice has consequences far beyond logos. It affects how customers trust each product, how search engines understand each company, how easily a company could be separated, and how much damage a problem in one brand can do to the others. This note explains the brand architecture we chose for Oryvelon, and why. The spectrum of options Brand strategists often describe a spectrum between two extremes. Branded house. One master brand, with products as descriptors: Parent Mail, Parent Maps, Parent Pay. Everything benefits from the parent's reputation — and everything shares its risks. Customers assume a single company, a single account and, often, a single data policy. House of brands. Independent brands, each with its own name and identity, owned by a parent that customers may barely notice. Each brand can target its audience precisely and stand on its own. Endorsed brands. Somewhere in between: independent brands with a visible endorsement — "by Parent" or "a Parent company". There is no universally right answer. The right choice depends on how similar the products and audiences are. What we chose, and why Oryvelon's companies serve very different people: merchants running Shopify stores (MerchNivo); brands and creative professionals (CastLyra); students, families and schools (EduRelia); consumers interested in astrology and readings (ZodiVela); people exploring visas and relocation (KeşifAtlası); shoppers buying accessories (Noveniq); brands looking for digital services (WeAreMedia); clients and students in beauty (Sinem Keser Beauty Academy). A single master brand across all of these would confuse customers and blur promises. A school does not want its learning platform branded like a horoscope app. A merchant does not want their operations tool to share a name with a talent marketplace. So we chose a house of brands with a light endorsement. Each company has its own name, identity, domain and voice. Oryvelon stays in the background as the parent, and the relationship is shown in a few natural places. How "An Oryvelon company" is used The endorsement is deliberately modest. Our rules are simple. Where it appears: - in the footer of a company's website, as a short line: An Oryvelon company; - on a company's about page, as a sentence such as MerchNivo is developed by Oryvelon; - on the company's profile on oryvelon.com's Companies page; - in press and partnership contexts where the group relationship is relevant. Where it does not appear: - as part of any company's logo; - larger or more prominent than the company's own brand; - as a link repeated on every page purely for search purposes; - in ways that erase the company's independent identity. The endorsement exists to add trust and clarity — so customers, partners and journalists can see who stands behind a company — not to build a network of backlinks or to make the group look bigger than it is. Why a quiet parent brand is an advantage A background parent brand has several practical benefits. Each company can speak directly to its customers. ZodiVela can be warm and personal. EduRelia can be careful and reassuring. MerchNivo can be practical and efficient. None of them has to compromise its voice to fit a group identity. Risk is contained. If one company has a difficult moment — a service outage, a negative review, a controversy in its market — the others are not automatically associated with it in customers' minds. Each company is separable. A company whose identity does not depend on the parent can take a partner, raise capital or be sold without rebranding. See Designing every company so it could stand alone. The parent can be precise. Oryvelon's own brand can focus on what it actually does — build and operate companies — rather than trying to represent every product. How we choose names Naming is step four of our new-company protocol, and it comes after the problem, the revenue mechanism and the recurring loop are clear. By then we know enough about the company to judge whether a name fits. We look for names that are: Distinctive. Easy to remember and hard to confuse with competitors. Invented or combined words often work better than generic descriptions. Pronounceable across languages. Our companies serve global and Turkish audiences. A name that works in English but is awkward in Turkish, or vice versa, creates friction. Suggestive without being literal. A good name hints at the category — commerce, talent, education, discovery — without locking the company into one feature. Available as a domain. Ideally the .com. For a market-specific company, a local domain may be the right primary choice; KeşifAtlası uses a Turkish name because it starts from Türkiye. Clear of conflicts. Checked against existing brands in the same category. Once a name is chosen, the domain is secured immediately: registered to the group, locked, set to auto-renew and managed under an account protected by two-factor authentication. See Domain, DNS and email security for a multi-brand group. Domains as part of the architecture Each company lives on its own domain, and that is not just branding — it is architecture. Separate domains mean: separate cookies and sessions, reinforcing data boundaries; separate email sending domains, each with its own authentication and reputation; separate search presence, so each company builds its own authority; clean separability if a company ever leaves the group. The group's own domain, oryvelon.com, hosts only the corporate site: the group's story, the company profiles, the build model and these insights. It does not host any company's product. Founder authority sites Beyond the parent and the companies, there is a third layer: the founders' own sites. ugurkeser.art and sinemkeser.com.tr are personal authority sites focused on each founder's expertise. They are not company sites, and they are not the parent brand. But they are naturally connected: a founder writing about e-commerce operations may point readers to MerchNivo; a founder writing about creators and fashion may point to CastLyra. Those links are contextual, measured with consistent UTM parameters and used only where they genuinely help the reader. See Founder authority sites and how they relate to a company group. Consistency for search engines and AI assistants Brand architecture now has a new audience: search engines and AI assistants trying to understand which entities exist and how they relate. When someone asks an assistant "Who owns MerchNivo?" or "What is Oryvelon?", the assistant draws on what it can find and trust. Consistent descriptions help: each company's own site describes the company clearly; oryvelon.com has a profile for each company using the same core description; structured data on oryvelon.com marks each company as part of the Oryvelon organisation; founder sites describe the founders' roles consistently. When all of these sources agree, it becomes easy for machines to understand the relationships correctly. When they contradict each other, confusion follows. We describe the technical side in Structured data and llms.txt. Mistakes we avoid Forcing a shared visual identity. Companies should look like themselves, not like siblings in uniform. Over-endorsing. A parent logo on every page of every product makes the parent the brand, which defeats the purpose. Link networks. Linking every site to every other site on every page for search purposes looks manipulative and helps nobody. Late domain decisions. Building a brand before securing its domain invites expensive problems. Inconsistent descriptions. Describing a company one way on its site and another way on oryvelon.com confuses customers and machines. Voice and positioning, company by company Independent brands need independent voices. Here is how each Oryvelon company's positioning shapes the way it speaks — the kind of guidance that lives in each company's source of truth. MerchNivo speaks like a capable operations colleague: practical, concise, focused on what the merchant should do next. It never overstates what AI can do, and numbers always come from the store. CastLyra speaks to two audiences at once. To brands it is professional and efficient; to talent it is respectful and protective. Verification and consent are part of its promise, so it talks about them openly. EduRelia is calm, careful and encouraging. It speaks to families and schools with clarity about privacy, and to students with warmth appropriate to their age. ZodiVela is warm, personal and a little poetic — while being clear that readings are for reflection and entertainment, not predictions or professional advice. KeşifAtlası is clear and trustworthy. It explains complex rules in plain language, is honest about uncertainty, and never promises outcomes that depend on authorities. Noveniq is a consumer commerce brand: product-focused, straightforward, helpful. WeAreMedia speaks like a senior agency partner: confident, commercially minded and accountable for results. Sinem Keser Beauty Academy is personal and expert, built around craft, care and education. Seen together, these voices would clash inside one master brand. As independent brands, each can be exactly what its customers need. When the parent should be more visible A light endorsement is the default, but there are moments when showing the group more clearly helps a company. Early trust for a new company. A company in its first months has no reviews or history. A clear statement that it is built and operated by an established group — with a link to the group's profile — can reassure early customers and partners. Institutional customers. Schools, larger brands and partners often want to know who stands behind a product before signing a contract. For EduRelia and CastLyra, the group relationship is relevant in sales conversations and documentation. Press and events. Journalists writing about one company often want to understand the wider group. The Press page provides a consistent description. Even in these cases, the company's own brand stays primary. The parent adds context; it does not replace identity. A checklist for choosing your own brand architecture If you run more than one product and are deciding how the brands should relate, these questions usually point to the right answer. Do your products serve the same customers in the same role? If yes, a branded house may work. If no, lean towards independent brands. Do they make different promises about data or behaviour? Different promises need different names, so customers are never confused about who holds what. Could any product one day be separated, sold or partnered? Independent brands make that far easier. Would a problem in one product damage the others if they shared a name? If yes, keep them apart. Does the parent have its own story worth telling? If so, give it a quiet, precise brand of its own — as Oryvelon has — rather than stretching it across every product. Can you keep descriptions consistent everywhere? Whatever you choose, inconsistency is the real enemy. Are all the domains secured? Do this before spending anything on design. Most multi-product groups serving different audiences end up where we did: independent brands, a quiet parent and a clear, modest endorsement. A naming process, step by step Naming is where brand architecture becomes practical. The process we follow is simple and repeatable. 1. Write the brief first. A paragraph from the company's source of truth: who the customer is, what the company promises, which markets it starts in and what tone it should have. Without this, naming turns into a contest of personal taste. 2. Generate widely. Invented words, combinations of meaningful roots, words from other languages. We deliberately produce many candidates before judging any of them. 3. Filter hard. Most candidates fail quickly: the domain is taken, the name means something unfortunate in Turkish or another target language, it is too close to an existing brand in the category, or it is hard to spell after hearing it once. 4. Say it out loud. We read the survivors aloud in English and Turkish, and imagine them in a sentence a customer would say: "I checked it on…", "Our store uses…". Names that look good but sound awkward fall away here. 5. Check the footprint. Before a final decision we look at the domain, the main social handles, and whether a clear trademark conflict exists in the relevant classes. Where there is real legal doubt, that is a question for a qualified adviser, not for a naming session. 6. Secure, then announce. The domain is registered before the name is shared widely. It is a small cost compared with the price of losing a name after design work has started. The brand footprint a new company needs A name on its own is not a brand. When a company is created, we set up a standard footprint so every company starts from the same, clean baseline. Element Standard Primary domain Registered to the group, locked, auto-renew on, 2FA on the registrar account Defensive domains Only where there is a clear risk of confusion or misuse — not dozens by default Email Separate sending domain with SPF, DKIM and DMARC before the first message goes out Support and contact addresses On the company's own domain, never a group address Social handles Reserved in the company's name, even if not used at launch Analytics The company's own property, with the group's UTM standards First website A focused phase-zero site that tests the promise before the product exists The point of the standard is not bureaucracy. It is that a company never has to untangle itself later from a shared inbox, a group analytics account or a social handle registered in someone's personal name. Worked example: a name that follows the market KeşifAtlası is the clearest example of a name shaped by its first market. The product starts from Türkiye, explaining visa and relocation rules for people there. A Turkish name — roughly "discovery atlas" — tells that audience immediately that the product speaks their language and understands their starting point. That choice has trade-offs, and we accept them knowingly. The name contains characters that are harder to type on some keyboards, so the domain uses the plain spelling, kesifatlasi.com, while the brand keeps its proper spelling everywhere it is displayed. If the company later serves other starting countries, the brand may need a companion name or localised pages. That is a future decision, recorded as an open question rather than solved in advance. Compare that with MerchNivo or CastLyra, which serve global audiences from the start. Their names are invented, easy to pronounce in many languages and carry no single-market meaning. Where the services company fits Brand architecture also decides where different kinds of work live. Oryvelon builds and operates companies; it does not sell agency services. Client work for other brands happens in WeAreMedia, under its own name and domain. Keeping that line clear protects both sides. Clients of WeAreMedia deal with a services company that is accountable for their results. Customers of the product companies never wonder whether the brand is really an agency. And oryvelon.com can describe the group honestly. We explain how that works in A services company inside a company group, and the product side of the same principle in Why we don't merge products into one app. Brand architecture and AI visibility A newer reason to get the architecture right is that people increasingly ask assistants rather than search engines. An assistant that finds MerchNivo described as "an AI e-commerce employee for Shopify stores" on its own site, on its oryvelon.com profile and on a founder's site will repeat that description with confidence. One that finds three different descriptions will hedge, or mix companies up. That makes the one-sentence description of each company a brand asset in its own right. We write it once, in the source of truth, and reuse it word for word in profiles, structured data and press material. More on this in Search and AI visibility for companies. Common naming mistakes Naming before the problem is clear. A name chosen for an idea that later changes shape rarely fits the company that ships. Describing a feature. Literal names like "store-report-bot" age badly as soon as the product grows beyond that feature. Testing only in one language. For a group serving Turkish and global audiences, a name must survive both. Choosing a name you have to explain. If it needs spelling out on every call, customers will struggle to find it again. Registering in a personal account. Domains and handles belong to the group from day one, so ownership is never in doubt. Summary Oryvelon's brand architecture is a house of brands with a light endorsement. Each company has its own name, domain, identity and voice, chosen carefully and secured early. The parent stays in the background, visible in footers, about pages, company profiles and press — present enough to add trust, quiet enough to let each company be itself. It is the brand equivalent of our technical rule: share the foundations, not the identity. ## The operating cadence behind a portfolio of digital businesses URL: https://oryvelon.com/insights/operating-cadence-for-a-portfolio/ Category: Company Building People often imagine that running a group of companies requires a large headquarters, many meetings and complicated reporting. It does not — or at least, it should not. What it requires is a rhythm: a predictable cadence of small, consistent reviews that makes problems visible early and turns decisions into habits. At Oryvelon the operating cadence is one of the quiet parts of the operating core. It is not glamorous. But it is the reason a small team can keep eight very different companies healthy at once. Why cadence matters more than strategy Strategy decides where to go. Cadence determines whether you notice when you are going somewhere else. Most problems in digital businesses start small: a slowly rising AI bill, a gradual drop in retention, a domain renewal nobody noticed, a former contractor who still has access, a tool that three companies pay for separately. None of these is dramatic on the day it starts. All of them become expensive if nobody looks. A regular cadence is simply a promise to look — at the right things, at the right frequency, in a consistent way. The layers of the cadence Our cadence has four layers, each with a different frequency and purpose. Frequency Purpose Main questions Continuous Alerts Is anything broken right now? Weekly Health signals Is each company behaving normally? Monthly Metrics and economics Is each company getting healthier? Quarterly Standards and hygiene Are access, tools, costs and standards in order? On top of these sit event-driven reviews: launch checkpoints at 30, 60 and 90 days, capital allocation reviews, and incident reviews when something goes wrong. Continuous: alerts Some things cannot wait for a weekly review. Automated alerts cover: sites or products that stop responding; spikes in errors; AI budgets crossing soft limits; failed background jobs; email delivery problems; certificate or domain expiry warnings. Alerts go to the people responsible for the affected company. The rule is that an alert must be actionable: if an alert fires often and nobody acts on it, it is either fixed or removed. Noise trains people to ignore alerts, which is worse than having none. Weekly: health signals Once a week, each company's owner looks at a short set of health signals. It takes minutes, not hours. traffic and sign-ups compared with the previous weeks; activation — did new users reach the first meaningful action?; any unusual changes in usage or revenue; open support issues and recurring complaints; AI and infrastructure spend compared with expectations; anything shipped, and anything that broke. The weekly look is not a report for anyone else. It is a habit that keeps each owner close to their company. If something looks wrong, it is investigated immediately rather than waiting for the monthly review. Monthly: metrics and economics per company Every month, each company produces a short snapshot. The template is identical across the group, which makes companies easy to compare. Key customer metrics against the targets set at launch — activation, retention, paying customers, repeat purchases, depending on the company. Economics — revenue, and costs broken down into hosting, database, AI, email, tools and a fair share of shared costs. See Unit economics per product. What changed — shipped improvements, experiments and their results. What customers said — a few real quotes, good and bad. Risks — anything that could hurt the company in the coming months. One request — the single thing the company most needs from the group. The monthly snapshot fits on one page. Writing it forces clarity; reading several side by side gives the group a picture of the whole portfolio. Quarterly: standards and hygiene Every quarter, the group runs a set of hygiene reviews across all companies. These are the checks that nobody enjoys and everybody benefits from. Access review Who has access to which system, for which company, with which role? Former contractors are removed, roles are tightened to what people actually need, and two-factor authentication is confirmed on every critical account. See Least-privilege access for small teams and contractors. Tool and subscription review Every tool and subscription is listed with its owner, its users, the companies it serves and its cost. Duplicates are closed, unused seats removed, and costs attributed to the right company. See The quarterly tool review. Cost review Per-company costs are reviewed for trends: AI spend growing faster than usage, infrastructure that could be right-sized, vendors whose pricing has changed. See Cost discipline for AI products. Domain, DNS and email review Domains are checked for renewal dates, locks and ownership. DNS records are compared against the standard. Email authentication — SPF, DKIM and DMARC — is verified on every sending domain, and DMARC reports are reviewed. See Domain, DNS and email security for a multi-brand group. Standards review The shared standards themselves are reviewed. Are they still right? Have companies found better approaches that should become the standard? Is anything outdated? Changes are made deliberately and communicated to every company. Event-driven reviews Some reviews happen when something triggers them rather than on a calendar. Launch checkpoints. Every company and major feature has written 30/60/90-day KPIs and a decision at each checkpoint. Allocation reviews. Resources are reallocated across companies based on evidence, usually aligned with checkpoints and monthly snapshots. Incident reviews. When something goes wrong — an outage, a data mistake, a security issue, a costly bug — there is a short, blameless review: what happened, why, how it was fixed and what changes so it is less likely anywhere in the group. The last part matters most: a lesson learned in one company becomes a standard for all of them. Keeping the cadence light A cadence can easily turn into bureaucracy: long meetings, elaborate dashboards, reports written to impress rather than inform. We protect against that with a few rules. Templates are short. A monthly snapshot is one page. A checkpoint review is two. Written first, discussed second. Reading a snapshot takes less time than presenting it. Meetings focus on decisions, not updates. Decisions are recorded. Every review ends with written decisions, or an explicit "no change". No status theatre. Nobody gets credit for long reports. Clarity and honesty are what count. Automate the collection. Numbers that can be pulled automatically are; people spend their time on interpretation. How the cadence helps each company The cadence is not only for the group's benefit. Each company gains from it too. Owners stay close to their numbers without building their own reporting. Problems surface early, while they are small. Security and hygiene are handled consistently, so no company is the weak link. Each company can see how it compares with others at similar stages. Requests to the group are heard regularly, not only when something is urgent. What a typical month looks like To make the rhythm concrete, here is how an ordinary month flows through the group. Week one. Monthly snapshots are written for each company and the core. The founders read them in one sitting, note questions and identify any company that needs a closer look. Decisions that can be made immediately are recorded. Week two. Follow-ups happen with the companies that raised requests or showed unusual numbers. If a company has reached a launch checkpoint, its review happens here. Week three. Build capacity is focused on the priorities agreed in week one. The core team works on standards improvements that benefit several companies. Week four. A lighter week for reviews, heavier for building. Weekly health checks continue as always. Anything unusual found during the month is written down for the next snapshot. In the last month of each quarter, the quarterly hygiene reviews are added: access, tools, costs, domains and standards. They are spread across the month rather than crammed into a single day. The cadence across different kinds of company The layers are the same for every company, but what each layer looks at depends on the business. MerchNivo: connected stores, briefings opened, actions taken, AI cost per store, background job reliability. CastLyra: verified profiles, active businesses, requests sent and answered, time to first response — on both sides of the marketplace. EduRelia: active students and classes, teacher engagement, school licence usage — with extra attention at the start and end of terms. ZodiVela: readings per user, return visits, credit and subscription behaviour, content safety signals. KeşifAtlası: tests completed, reports purchased, rule updates published, questions routed to people. Noveniq: orders, repeat purchases, stock health, returns. WeAreMedia: client retention, delivery quality, margins per engagement. Sinem Keser Beauty Academy: appointments, course enrolments, repeat clients and students, digital product sales. Because the template is shared, the group can compare companies' health without pretending they are the same kind of business. The tools behind the cadence We keep the tooling for the cadence simple. Dashboards per company, fed automatically from each company's analytics, billing and infrastructure — never a combined dashboard that mixes customer-level data across companies. A cost view per company, combining invoices from hosting, databases, AI providers, email and tools, with shared costs allocated transparently. A shared document space where snapshots, checkpoint reviews and decisions are written and kept. Alerting routed to the right owner for each company. Scheduled automation for routine checks — site health, search visibility, certificate and domain expiry — with a short summary delivered to the people responsible. None of this requires an elaborate internal platform. It requires consistency, and the discipline to keep looking. Signs the cadence is working — or not A cadence should be judged by its effects, not by how many reviews happen. These are the signals we watch. Signs it is working: - problems are usually found by our own reviews before customers report them; - renewals, certificates and access changes never cause surprises; - monthly snapshots get shorter and clearer over time, because owners know what matters; - decisions are traceable — anyone can find out why a company received more or less support; - lessons from one company's incident appear as standards in others within weeks. Signs it is not: - reviews happen but nothing changes because of them; - the same risk appears in several monthly snapshots without action; - owners write snapshots for the group rather than for themselves; - alerts are ignored because too many are noise; - hygiene reviews are postponed "until things calm down". When the second list starts to appear, the fix is rarely more process. It is usually less: fewer metrics, shorter templates, fewer alerts, and a renewed commitment to act on what the reviews show. A weekly check that catches something The value of the weekly look is easiest to see with an example. Imagine the owner of MerchNivo doing the ten-minute review on a Monday. Sign-ups and connected stores look normal. Briefings opened are steady. But AI spend is noticeably higher than the week before, while the number of briefings delivered has not changed. That combination is the signal: cost per useful outcome has gone up. The owner opens the gateway metrics for the week. Validation failures on the briefing feature have risen, so many requests are being retried. A change to how order data is formatted, shipped the previous Thursday, is sending the model a field it handles badly. The fix is small. Without the weekly look, it would have surfaced at the end of the month as an unexplained line in the cost report. Nothing in that story needed a meeting, a dashboard project or a new process. It needed one person looking at a short list of numbers at the same time every week. The kinds of store-level signals MerchNivo itself watches for merchants are described in Operations signals for Shopify stores — the same idea applied to our own operations. The monthly snapshot, filled in A template is easier to follow when you can see one filled in. Here is an illustrative snapshot for Noveniq, with bracketed placeholders where real figures would go. Noveniq — monthly snapshot Metrics vs targets. Orders [n] vs target [n]. Repeat purchase rate [x%] vs [y%]. Return rate [x%], higher than last month. Economics. Revenue [amount]. Costs: hosting [amount], apps and tools [amount], email [amount], share of shared costs [amount]. Margin per order [amount]. What changed. New bundle page launched mid-month; early add-to-cart rate looks better, too early to judge. What customers said. Two returns cited a cable length that was not clear on the product page. Risks. One supplier has lengthened its lead time; two fast-moving products could go out of stock next month. One request. Help updating product pages with clearer specifications. It takes perhaps half an hour to write. It tells the reader what happened, why returns rose, what could go wrong next and what help is needed. Stock accuracy in particular gets a dedicated line for commerce companies; we explain why in Inventory truth. How the cadence changes with a company's stage The layers stay the same, but their weight shifts as a company matures. Stage Where attention goes Before launch Weekly: progress against the build plan and phase-zero signals. Monthly snapshot is short: what was learned, what it cost. First 90 days Weekly checks carry the most weight. Checkpoints at 30, 60 and 90 days drive the key decisions. Steady operation The monthly snapshot and quarterly hygiene do most of the work. Weekly checks stay brief. Scaling More detailed economics, closer cost monitoring, extra attention to access as more people join. Winding down A short plan: customer communication, data export and deletion, domain and tool cancellations, access removal. The last row deserves a word. Stopping a company is part of the cadence too, and it has its own checklist. Customers are told clearly and early. Data is returned or deleted according to the company's rules. Subscriptions are cancelled and access is removed, so a closed company does not leave open doors behind it. Common mistakes with an operating cadence Measuring everything. A snapshot with thirty metrics hides the three that matter. We keep each company's list short and change it deliberately. Reviewing without deciding. A review that ends without a written decision, even "no change", tends to be repeated next month with the same conclusion. One template for the wrong purpose. The monthly snapshot is for owners and founders. Using the same document to impress a partner turns it into marketing, and honesty suffers. Skipping hygiene when busy. Access and secrets reviews are the first to be postponed and the most expensive to forget. A former contractor's key that still works is not a hypothetical risk. See Secrets management basics. Combining data to compare companies. Comparison works on aggregated metrics per company. It never requires pooling customer-level data from different companies into one place. Starting a cadence in your own business You do not need several companies to benefit from this. A single product team can start with three habits: A fixed weekly slot, fifteen minutes, looking at five or six numbers that describe the business's health. A one-page monthly note: metrics against targets, costs, what changed, what customers said, one risk and one request. A quarterly list: who has access to what, which tools are paid for, which domains and certificates renew soon. Run it for a quarter before adding anything. Most teams find that the discipline of looking regularly matters far more than the sophistication of what they look at. Summary The operating cadence is the unglamorous heartbeat of Oryvelon: continuous alerts, weekly health signals, monthly metrics and economics, quarterly hygiene reviews and event-driven decisions. It is light by design and consistent across every company. It is how a small group keeps several very different businesses healthy — by looking at the right things, regularly, and writing down what it decides. ## What an AI gateway does — and what it should never do URL: https://oryvelon.com/insights/building-an-ai-gateway/ Category: AI Operations When several products use AI, a question arrives quickly: should each product call model providers directly, or should calls go through something shared? Calling directly is simplest at first. Each team picks a model, adds an API key and writes some code. But after a few products, the same problems appear everywhere: no consistent budgets, no shared monitoring, no standard way to handle failures, keys scattered across projects, prompts edited in place without history, and no easy way to switch models when prices or quality change. At Oryvelon, every AI feature goes through the Oryvelon AI Gateway: a thin shared layer that standardises how products use models. This note explains what it does, how it is designed, and the line it must never cross. The gateway in one sentence The AI gateway is a common, product-aware layer between our applications and model providers that applies each product's routing, budgets, limits, safety checks, logging and fallbacks — without ever mixing one product's prompts, knowledge or users with another's. What the gateway is responsible for Request policy Each product has a policy: which model families it may use, which features may call which models, maximum input and output sizes, and rate limits per user and per product. The gateway enforces the policy on every request. A consumer feature cannot accidentally call an expensive reasoning model; an internal batch job cannot flood a provider. Model routing Simple tasks such as classification or short summaries go to smaller, cheaper models. Complex reasoning goes to stronger models, sparingly. The routing rules live in configuration, per product and per feature. See Model routing: matching tasks to the right AI model. Prompt registry Important prompts are registered with an identifier, a version, an owner and a test status. Applications reference prompts by identifier and version rather than embedding text in code. That makes changes visible, reversible and testable. See Prompt versioning and evaluation for production AI. Budget controls Every product has its own monthly budget with a soft limit (alert) and a hard limit (switch to degraded mode). Budgets can also be set per feature or per user where abuse is possible. See Cost discipline for AI products. Safety hooks Before a request leaves, the gateway can apply product-specific checks — for example, removing personal information that the model does not need. After a response returns, it can run output checks: content boundaries for consumer products, format checks for structured outputs. See Safety boundaries for consumer AI products. Observability For every request the gateway records latency, token usage, cost, model, prompt version, success or failure, and retries — per product. Logs are filtered so that sensitive raw content is not stored by default. Schema validation Where a product asks for structured output, the gateway validates the response against the expected schema before returning it. Invalid responses are retried within limits or reported as failures, so application code never receives malformed data. See Structured outputs. Fallbacks If the primary model fails — timeout, provider outage, repeated invalid output — the gateway follows the product's approved fallback: a secondary model, a simpler prompt, or a signal to the application to use its non-AI degraded mode. See Fallbacks and degraded modes for AI features. Caching For deterministic, low-variance requests, the gateway can return cached responses. Cache keys include the product, the prompt version and a fingerprint of the input, so one user's output is never served to another and cached answers expire when inputs change. See Latency, caching and batching. What the gateway must never do This part matters more than the list above. A shared AI layer is exactly the kind of component that can quietly break the boundaries between companies. So the gateway has explicit rules about what it must not do. It never merges prompts. Each product's prompts are registered under that product. There is no "shared system prompt" that every product uses. It never shares knowledge bases. If a product uses retrieval over its own documents, that index belongs to the product and is only queried on its behalf. See Retrieval without leakage. It never builds user profiles. The gateway does not know, or try to know, that the same person uses two products. It sees requests from products, not people across products. It never pools keys or budgets. Each product has its own AI project and API keys with the provider, and its own budget. The gateway routes a product's requests using that product's credentials. It never becomes the source of truth. The gateway does not store business facts and does not make business decisions. Facts come from each product's verified systems; decisions come from each product's rules. See AI explains, verified data decides. It never logs sensitive content by default. Operational metrics are recorded; raw inputs and outputs are only kept where a product explicitly needs them for evaluation, with personal information filtered and retention limited. How it is designed The gateway is intentionally thin. It is a small service plus a client library that each product uses. Most of its behaviour is configuration: policies, routes, prompt references, budgets and fallback chains, stored per product. A typical request flows like this: The product calls the client with a feature name, a prompt identifier and version, and the input data (which already contains the verified facts). The gateway loads the product's policy for that feature and checks limits and budget. Pre-request safety checks run. The request is routed to the configured model using the product's credentials. The response is validated, post-response checks run, and metrics are recorded. On failure, the configured fallback runs. The product receives either a valid response or a clear failure signal. Built when needed, not before It is easy to over-engineer something like this — to build a large internal AI platform with every feature imaginable before any product needs it. That is what our master plan calls "architecture theatre", and we deliberately avoid it. The gateway grew in steps. Budgets and logging came first, because every product needed them from day one. Routing followed when products began to mix small and large models. The prompt registry arrived when prompts became numerous enough that editing them in code was risky. Caching was added for the products where it clearly paid off. Each capability exists because a real product needed it. What it gives each product For a product team, using the gateway means: cost limits and monitoring from the first request; a standard way to switch or combine models; versioned prompts and a place to run evaluations; consistent fallbacks when providers fail; safety checks appropriate to the product; no need to handle provider credentials in application code. And it means all of that without giving up the product's independence. If a product were ever separated from the group, it could keep running the client with its own configuration, or replace it — its prompts, keys and data were always its own. See Designing every company so it could stand alone. A worked example: one request, end to end Abstract descriptions only go so far, so here is what happens when a merchant opens their morning briefing in MerchNivo. The briefing feature asks the gateway for a short, structured summary of yesterday's store activity. Before the call is made, MerchNivo's own code has already pulled the numbers from the store: orders, revenue, refunds, low-stock products, abandoned checkouts. Those figures are the facts. The model's job is only to put them into plain language and suggest which two or three things deserve attention first. The request reaches the gateway carrying four things: the product identifier (MerchNivo), the feature name (daily_briefing), a prompt reference (briefing.summary, version 7) and the input payload with the verified numbers. The gateway looks up MerchNivo's policy for daily_briefing. The policy says this feature may use a mid-sized model, has an output limit suitable for a short summary, and must return JSON matching the briefing schema. It checks that MerchNivo's monthly budget has room and that this store has not exceeded its daily allowance of briefings. A pre-request hook strips anything the model does not need — customer names and emails that happen to appear in order data, for example. The model gets "3 refunds today, two for the same product", not who asked for them. The call goes out using MerchNivo's own credentials. The response comes back, and the gateway validates it against the schema: a headline, a list of up to three priorities, each with a reference to the underlying figure. One of the priorities refers to a product that was not in the input. That is a hallucinated reference, so the validator rejects the response and the gateway retries once with the same prompt. The second response is valid. Metrics are recorded: model, prompt version, tokens, cost, latency, one retry, success. No raw content is stored. MerchNivo receives a clean, validated object and renders it in the merchant's dashboard. If both attempts had failed, the gateway would have returned a clear failure, and MerchNivo would have shown its non-AI fallback: the same numbers in a simple table, without the narrative. The merchant still gets the facts. They just get them without the prose. Configuration per company Because the same gateway serves very different companies, most of what makes it useful is configuration rather than code. A few examples show how differently it is set up. ZodiVela is a consumer product with a large number of short interactions. Its configuration emphasises per-user rate limits, strict content boundaries on output, small models for most tasks and aggressive budget alerts. Readings are framed for reflection and entertainment, and the post-response hooks check that the language stays within those limits. KeşifAtlası uses AI to explain rules, never to decide them. Its configuration requires that every explanation references rule identifiers from its verified rules database, and the validator rejects responses that make eligibility claims without such a reference. Questions the system cannot answer confidently are routed to a person. See Rules engines for eligibility. EduRelia handles content for students. Its configuration minimises what is sent — no names, no identifiers beyond what the task requires — and applies age-appropriate output checks. Nothing from EduRelia's requests is retained for evaluation without explicit, documented need. See Data minimisation for children in edtech. CastLyra uses AI to help structure profiles and briefs. Its configuration keeps talent contact details out of prompts entirely and validates that generated profile text does not invent credentials. Same gateway, four very different sets of rules. That is the point: the shared layer handles the mechanics, and each company decides what is appropriate for its own customers. What we monitor The gateway's metrics are only useful if someone looks at them. Each company's owner sees a small set of indicators for their own product, as part of the weekly health check described in The operating cadence behind a portfolio of digital businesses. Cost per useful outcome. Not cost per call, but cost per briefing delivered, per reading completed, per explanation shown. This is the number that tells us whether AI is paying for itself. Validation failure rate. A rising rate usually means a prompt has drifted, an input format has changed or a model update has altered behaviour. Fallback rate. How often users receive the degraded experience. A small number is healthy; a growing one needs attention. Latency at the slow end. Averages hide the experiences that frustrate people. We watch the slowest requests. Budget position. Where each product stands against its soft and hard limits for the month. Alerts fire on sudden changes in any of these. Slow trends are discussed in the monthly snapshot. Evolving the gateway without breaking products Shared infrastructure has a particular risk: a change made for one product can surprise another. We handle gateway changes with the same discipline we apply to shared code in general. The client library is versioned, and each product upgrades on its own schedule. New capabilities are added behind configuration flags that default to off, so a product only gets new behaviour when it asks for it. Changes to default behaviour — for example, a stricter logging filter — are announced to every product owner before they take effect, with a date and a way to test in advance. Model changes are handled per product, not globally. When a provider releases a new model, one product evaluates it against its own test set first. If it performs better, that product switches. Others follow when their own evaluations agree. No product is moved to a new model because another product preferred it. Build, buy or skip: the trade-offs A gateway is not the right answer for every team. The choice depends on how many products use AI and how different they are. Situation Reasonable choice One product, one model, early stage Call the provider directly through a small internal wrapper. Add budgets and logging from day one. One product, several models or features A thin internal layer for routing, budgets and validation. Several products with different rules A shared gateway with per-product configuration, as we run. Strict regulatory or contractual requirements Whatever gives the clearest control over data flow and logging, even if it costs more to maintain. Off-the-shelf gateway software and provider-side features can cover a lot of the mechanics: routing, rate limits, usage tracking. What they cannot decide for you is the policy — which product may send what, what must be filtered, what counts as a valid answer. That part is always yours. Our own layer is mostly that policy, expressed as configuration. The wrapper in the first row matters more than it looks. Even a single product benefits from having every model call go through one function in its own codebase. When a second product appears, that function is where the gateway client slots in. Common mistakes when building an AI gateway Putting business logic in the gateway. It starts small: a formatting rule here, a pricing check there. Soon the gateway knows things about each product that only the product should know. Business rules belong in the product. One shared system prompt. It is tempting to add a group-wide preamble to every request "for consistency". It mixes products' voices and boundaries, and a change for one product silently changes all of them. Logging everything, just in case. Full request and response logs are convenient for debugging and a liability for everything else. Default to metrics; keep raw content only with a named purpose, filtering and a retention limit. A single budget for all products. It hides which product is expensive and lets one product's spike starve the others. Budgets are per product, and often per feature. Treating the gateway as a proxy only. A gateway that forwards requests without validating responses moves the hardest part — deciding whether an answer is usable — back into every application. Global model switches. Moving every product to a new model at once because it looked better in one product's tests. Each product evaluates for itself. A readiness checklist for a new AI feature Before any Oryvelon product ships a feature that calls a model, its configuration answers these questions: Which prompt identifier and version does it use, and who owns it? Which models may it call, with which limits on input and output size? What schema must the response match, and what happens when it does not? Which facts in the input come from verified systems, and which fields are removed before the call? What are the per-user and per-feature limits? What is the fallback, and what does the user see in degraded mode? What is logged, what is not, and for how long? Which metric tells us the feature is worth its cost? If any answer is missing, the feature is not ready, however good the demo looks. For a store-facing feature in Noveniq or a reading in ZodiVela, the answers differ completely — but the questions are the same. Where keys and environments fit Credentials deserve a short note of their own. Each product's provider keys live in a secrets store, scoped to that product and environment. Application code never sees them; the gateway retrieves the right key for the calling product at request time. Development and preview environments use separate keys with small budgets, and they never receive production customer data as prompt input. Test sets for evaluation are built from synthetic or properly anonymised examples. A mistake in a preview build should cost a few cents and expose nothing. Summary An AI gateway is a useful piece of shared infrastructure when it standardises how models are called and stays out of what products say to them. Ours handles routing, budgets, prompts, safety, observability, validation, fallbacks and caching for every Oryvelon product — and draws a hard line at merging prompts, knowledge, keys or users. Facts stay with each product's verified systems. Decisions stay with each product's rules. The gateway just makes calling a model safer, cheaper and easier to change. ## Model routing: matching tasks to the right AI model URL: https://oryvelon.com/insights/model-routing/ Category: AI Operations Most AI features do not need the biggest model available. A support message needs a category. An order note needs three fields pulled out of it. A daily summary needs a few numbers turned into sentences. These are small jobs, and paying large-model prices for them is one of the quickest ways an AI product loses its margin. Model routing is the discipline of matching each task to the right AI model: small, fast models for the routine work, larger models for the problems that actually need reasoning, and a clear rule for every feature saying which is which. At Oryvelon, routing is one of the first things our shared AI gateway does for every product. This article explains how we think about it, how we write the rules, and the mistakes we try to avoid. Why one model for everything is the expensive default When a team builds its first AI feature, it usually picks one capable model and uses it everywhere. That is a reasonable way to start. It removes a decision while the product is still finding its shape. The problem is that the default hardens. Six months later the same large model is classifying tickets, generating subject lines, extracting addresses and writing long explanations. Most of those calls are paying for capacity they never use. Latency is higher than it needs to be. And because nobody ever measured a smaller model on the simple tasks, nobody knows how much is being wasted. There is a quality argument too, and it surprises people. Larger models are not always better at narrow tasks. For a tight extraction job with a strict schema, a small model with a clear prompt can be just as accurate and more consistent, because there is less room for it to be creative. A model that likes to elaborate is a liability when you want exactly three fields and nothing else. So the question is never "which model is best?" It is "which model is good enough for this specific task, at the lowest cost and latency?" That question has a different answer for every feature. The task, not the product, is the unit of routing A common mistake is to route by product: "ZodiVela uses model A, MerchNivo uses model B." Products are bundles of very different tasks. Routing has to happen one level down. We group tasks into a handful of types, because each type has a predictable profile: Task type Typical example Usual model tier What matters most Classification Tag a support message by topic and urgency Small Consistency, speed Extraction Pull dates, amounts and names from a document Small to mid Schema adherence Short generation Write a two-line product highlight Small to mid Tone, length control Summarisation Turn a day of store activity into a briefing Mid Faithfulness to input Explanation Explain why a visa rule applies to a case Mid to large Accuracy, citing sources Multi-step reasoning Compare several options against constraints Large Correctness The tiers are starting points, not laws. Plenty of summarisation works well on small models; some classification tasks with subtle categories need more. But the table gives a team a sensible first guess, and it forces the right conversation: what kind of task is this, really? Once a feature is described as a task type, the routing rule almost writes itself. How we write a routing rule Every AI feature in an Oryvelon company has a routing entry in that product's gateway configuration. It is short, and it always has the same parts: Feature name. The stable identifier the application uses, such as daily_briefing or ticket_triage. Task type. One of the types above. Primary model. The model family and tier the feature uses by default. Fallback. What happens if the primary fails: a second model, a simpler prompt, or a signal to use the non-AI degraded mode. Limits. Maximum input and output size for this feature. Owner. The person responsible for the rule. Reason. One or two sentences explaining why this model was chosen, with a reference to the evaluation that supports it. The last field is the one teams skip, and it is the one that matters most a year later. "Mid-tier because the small model missed 1 in 8 refund-related priorities in the briefing test set" is a reason someone can re-check when a new model is released. "Seemed better" is not. A rule might look like this in plain terms: ticket_triage is a classification task. Primary: small model. Fallback: a different small model, then rule-based keyword tagging. Max input: 2,000 tokens. Owner: product lead. Reason: small model matched the labelled test set at the agreed threshold; mid-tier gave no meaningful improvement. That is all a routing rule needs to be. It is readable by anyone on the team, not only the engineer who wrote it. Start small and earn your way up Our default is to begin every feature on the smallest model that could plausibly do the job, and to move up only when evaluation proves a gap. This is the opposite of how most teams work. They start large, get a good result, and never test whether they needed it. We prefer the opposite order for three reasons. First, it forces a test set to exist. You cannot decide a small model is "not good enough" without examples of what good enough means. That test set is then reused for every future model decision, and it becomes part of the prompt's record in the registry described in Prompt versioning and evaluation for production AI. Second, it makes the cost of quality visible. If the mid-tier model scores slightly higher and costs several times as much, that is now a choice someone has made on purpose, with numbers in front of them. Third, it keeps the product honest about what the AI is for. When a feature seems to need a very large model, that is often a sign the task is badly shaped: the prompt is asking the model to find facts it should have been given, or to make a decision that should come from rules. Moving the facts and the decision out of the model frequently makes a small model sufficient. This is the practical side of our principle that AI explains, verified data decides. Measuring the trade-off honestly The cost side of routing is easy to see. The quality side takes more work, and it is where routing decisions usually go wrong. We look at four numbers for each candidate model on a feature's test set: Pass rate against the feature's quality bar. Defined per feature. For extraction, exact field matches. For summaries, a rubric that checks faithfulness to the input and coverage of the required points. Validation failure rate. How often the output fails the schema or reference checks described in Structured outputs: validating AI responses before users see them. Latency at the slow end. Not the average. The slowest tenth of requests is what users remember. Cost per useful outcome. Total spend, including retries and fallbacks, divided by the number of outputs that were actually delivered to a user. The fourth number changes decisions. Imagine a small model that costs a fraction of a mid-tier model per call but fails validation often enough to need retries, and occasionally falls through to the fallback. Its cost per delivered result can end up close to the mid-tier model's, with worse latency. On paper it was cheaper. In practice it was not. The reverse happens too. A large model with a great pass rate may be only marginally better than a mid-tier one on the cases users notice. If the gap sits in edge cases that the product handles elsewhere — for example by escalating to a person — the extra spend buys very little. We cover the budget side of this more fully in Cost discipline for AI products. Fixed rules first, dynamic routing later There are two broad styles of routing. Fixed routing sends every request for a feature to the same model. It is predictable, easy to audit and easy to reason about when something breaks. If a user reports a bad briefing, you know which model produced it. Dynamic routing decides per request. A cheap classifier, or a simple rule, looks at the input and chooses a tier: short, familiar inputs go to the small model; long, unusual or ambiguous ones go to a larger model. Done well, it gives large-model quality on hard cases at small-model cost on the rest. We start every feature on fixed routing. Dynamic routing earns its place only when three things are true: the feature's inputs vary clearly in difficulty; the difference can be detected cheaply and reliably before the main call; the test set shows the small model failing mostly on the hard inputs, not randomly. If the small model's failures are scattered across easy and hard cases alike, a difficulty router will not help. You will just add a second model call and a new place for bugs. When we do use dynamic routing, the router's rules are as explicit as possible. Input length, the presence of certain fields, the number of items to compare, a language flag. A second model deciding which model to call is sometimes justified, but it doubles the surface you have to evaluate, so we use it sparingly. Escalation within a request A related pattern is escalation on failure. The small model tries first. If its output fails validation, or a confidence check falls below a threshold, the request is retried on a larger model before anything is shown to the user. This works well for tasks where the small model succeeds most of the time and failures are detectable. Extraction with a strict schema is the classic case: a missing required field or an invalid reference is easy to spot, and the retry is cheap relative to the savings on all the calls that succeeded first time. It works badly where failure is not detectable. If a summary is fluent but subtly wrong, validation will pass it and the escalation will never trigger. For those tasks, escalation-on-failure gives a false sense of safety, and it is better to pick the right tier up front. Escalation between models is not the same as escalation to a person. When the question itself is too hard or too consequential for any model, the right route is a human, and that is designed separately. See Designing human escalation paths in AI products. How routing looks across our companies The same gateway serves very different products, so the routing tables look quite different from company to company. MerchNivo, the AI e-commerce employee for Shopify stores, runs many small tasks and a few larger ones. Tagging order notes, classifying customer messages and flagging unusual refunds are small-model work. The daily briefing, which has to weigh several signals and pick what deserves a merchant's attention first, sits on a mid-tier model. Imagine a store sees a spike in refunds on a single product overnight. Detecting the spike is arithmetic done in code, from the store's own data. Classifying the refund reasons is a small-model task. Writing the one-paragraph explanation that connects the spike to a recent supplier change mentioned in order notes is where a mid-tier model earns its cost. None of those steps asks a model to invent a number. ZodiVela is a consumer product with many short interactions. Most of its tasks run on small models with strict length and content limits. Longer readings use a mid-tier model, and per-user rate limits keep usage predictable. The value of routing here is mainly cost: in a consumer product with credits and subscriptions, the difference between tiers shows up directly in the product's unit economics. KeşifAtlası routes differently because accuracy matters more than speed. The eligibility decision itself never touches a model; it comes from the verified rules database. Explaining which rules apply to a person's case, in plain language and with references to rule identifiers, runs on a stronger model, because a weak explanation of a correct decision still misleads people. Simple tasks like normalising a free-text country name run small. See Rules engines for eligibility. EduRelia adds another constraint: what can be sent at all. Routing rules there sit alongside data minimisation rules, so a feature's model choice is made after deciding which fields may leave the product. See Data minimisation for children in edtech. Four products, four very different tables. The shared gateway handles the mechanics. Each company owns its own rules. When a new model is released Model providers release new models often, and prices change. A routing setup that cannot absorb that churn will either fall behind or lurch from one model to the next based on excitement. Our process is deliberately boring: One product with a relevant feature runs the new model against that feature's existing test set. The four numbers are compared against the current primary. If the new model is better on quality, cost or latency without losing on the others, the product owner changes the routing rule and updates the reason field. The change is rolled out to a share of traffic first, with validation and fallback rates watched closely. Other products evaluate independently. No product switches because another did. Because routing is configuration, a switch is a change to one entry, not a code release. Because every rule has a test set, the evaluation takes hours rather than weeks. That is what makes model choice a routine operational decision instead of a project. Common mistakes A few patterns show up again and again when teams set up routing. Routing by product instead of by task. It hides the easy savings inside every product. No test set. Without one, every routing decision is a guess and every model switch is a gamble. Chasing the newest model everywhere. New is not the same as better for your task. Evaluate first. Ignoring retries in the cost. Cost per call flatters cheap models that fail often. Letting the model fetch or decide facts. It pushes you toward larger models to compensate for a design problem. Dynamic routing too early. It adds a moving part before you know whether the inputs actually vary in difficulty. No fallback. Every route needs a plan for when the model is slow, down or wrong. See Fallbacks and degraded modes for AI features. No owner. A rule nobody owns is a rule nobody revisits. A short checklist for a new feature When a team adds an AI feature to one of our companies, the routing part of the review is a few questions: What task type is this? What facts does the model receive, and where do they come from? What does a good output look like, and is there a test set of at least a few dozen real-shaped examples? What is the smallest model that passes it? What does it cost per useful outcome, including retries? What is the fallback, and what does the user see if everything fails? Who owns the rule, and what is the reason written next to it? If the team can answer those, the routing rule is ready. If they cannot, the feature is not ready either, whatever model it uses. Caching can reduce the load further for repeated, low-variance requests, which we cover in Latency, caching and batching. Summary Model routing means matching each AI task to the smallest model that does it well, and writing that choice down as a rule with an owner, a fallback and a reason. We route by task type rather than by product, start every feature on a small model and move up only when a test set shows a real gap, and judge models by cost per useful outcome rather than cost per call. Fixed rules come first; dynamic routing and in-request escalation are added only where inputs genuinely vary and failures are detectable. Kept in configuration inside a shared gateway, routing lets each Oryvelon company absorb new models and price changes as routine decisions, while facts and decisions stay with each product's verified systems. ## Prompt versioning and evaluation for production AI URL: https://oryvelon.com/insights/prompt-versioning-and-evaluation/ Category: AI Operations A prompt looks like text, so teams treat it like text. Someone opens a file, rewrites a sentence to fix a complaint, deploys, and moves on. The complaint goes away. Two other behaviours change quietly, and nobody notices for a week. In production, a prompt is closer to code than to copy. It changes what the product does for every user at once, and its effects are hard to predict by reading it. That is why prompt versioning and evaluation is one of the standing practices across Oryvelon's AI products: every important prompt lives in a registry, has numbered versions, is tested against a fixed evaluation set before release, and can be rolled back in minutes. This article describes how that works and why each part exists. Why prompts drift when nobody is watching Prompt problems rarely come from one bad edit. They come from many reasonable ones. A support lead asks for a friendlier tone. An engineer adds a line to stop the model mentioning a discontinued feature. A product manager adds an example to fix one customer's case. Each change is sensible. Together they produce a prompt that is twice as long, contradicts itself in two places and behaves differently from the version everyone thinks is live. Three things make this worse with AI than with ordinary code. First, the effect of a change is not local: adding a sentence about tone can change how the model formats numbers. Second, the model underneath may change too, so the same prompt can behave differently next month. Third, there is usually no test that fails. Ordinary code breaks loudly. A drifting prompt just gets slightly worse. Versioning and evaluation exist to make those changes visible and deliberate. What a prompt registry is A prompt registry is a central record of every prompt a product uses in production. Our AI gateway holds one per product. Products never share a registry, and there is no group-wide system prompt that every company inherits. Each registered prompt has: An identifier. A stable name such as briefing.summary or rules.explain. Versions. Numbered, immutable. Version 7 is always exactly version 7. An owner. The person who approves changes. A changelog entry per version. What changed and why, in a sentence or two. A test status. Which evaluation set it was run against, when, and the result. Its model binding. Which routing rule it is designed for, since a prompt tuned for one model tier may behave differently on another. See Model routing: matching tasks to the right AI model. Its output contract. The schema the output must match, if any. Applications do not contain prompt text. They ask the gateway for briefing.summary at a given version, or at whatever version is marked live for that feature. The registry is where the text lives. This one change — moving prompts out of application code and referring to them by identifier — does most of the work. It makes every prompt change a visible event with a name, a number and an author. Versions are immutable The single most useful rule in our registry is that a published version is never edited. If version 7 needs a fix, the fix is version 8. It sounds pedantic until the first time something goes wrong. If versions can be edited in place, "roll back to version 7" is meaningless, because version 7 is no longer what it was. Logs that record "prompt version 7" stop telling you what the model actually received. Evaluations run last month cannot be compared with evaluations run today. With immutable versions, every log line, every evaluation result and every incident report points at an exact text. That is what makes the rest of the system trustworthy. We apply the same rule to anything the prompt depends on at runtime. If a prompt includes a few fixed examples, those examples are part of the version. If it includes a list of allowed categories, that list is part of the version. Changing any of them creates a new version. What goes into an evaluation set A prompt without an evaluation set is a prompt you can only judge by reading. That is not enough. An evaluation set is a fixed collection of inputs for a feature, with a description of what a good output looks like for each one. We build them from four sources: Typical cases. Real-shaped inputs that represent everyday use. For a store briefing, an ordinary day with a normal mix of orders and a couple of refunds. Edge cases. The inputs that are unusual but legitimate. A day with zero orders. A store with a single product. A briefing input in which every figure is flat. Known failures. Every time a prompt misbehaves in production, the input (with personal information removed) becomes a test case. The set grows from real incidents. Adversarial cases. Inputs designed to push the model out of bounds: instructions hidden in user text, requests for things the product does not do, inputs that invite the model to invent a reference. Test inputs are synthetic or anonymised. We do not copy real customer records into evaluation sets, and for products involving children we do not use real student content at all. See Environments and test data and Data minimisation for children in edtech. For many features, a few dozen well-chosen cases is a useful starting point. Quantity matters less than coverage. Fifty variations of the same ordinary day tell you less than fifteen cases that each test something different. How outputs are scored Scoring is where evaluation becomes useful or useless. A vague rubric like "is it good?" gives vague answers. We break quality into checks that are as mechanical as possible. Automatic checks run on every case: Does the output match the schema? Does every reference in the output point to something that exists in the input? See Structured outputs: validating AI responses before users see them. Are required points present? For a briefing, is the largest change in the input mentioned? Are forbidden things absent? Invented numbers, promises the product cannot keep, advice outside the product's boundaries. Is the length within limits? Rubric checks need judgement: is the explanation clear, is the tone right for the product, does it read naturally. These can be scored by a person, or by a second model call with a strict rubric. When we use a model as a grader, we treat it as a tool that needs its own checks: we compare its scores with a person's on a sample before trusting it, and we never let the same prompt grade its own output. The result is a small table for each version: pass rates per check, with the failures listed. That table is what gets attached to the version's test status in the registry. The release gate A new prompt version goes live only when it passes a simple gate: It has been run against the feature's full evaluation set. It is at least as good as the current live version on every check marked as critical — schema validity, reference integrity, forbidden content. It improves something the change was meant to improve, and the improvement is visible in the results. Any regression on non-critical checks is understood and accepted in writing by the owner. The changelog says what changed and why. The second point is the one that prevents most incidents. A change that makes summaries more readable but occasionally lets an invented figure through is not an improvement. It is a regression with good prose. After passing the gate, the new version is rolled out to a share of traffic first. The gateway records the prompt version on every request, so validation failure rates, fallback rates and user feedback can be compared between old and new versions directly. If the numbers hold, the new version becomes live for everyone. Rollback in minutes, not days Every release plan needs a way back. For prompts, ours is a single configuration change: mark the previous version as live. Because versions are immutable and applications refer to prompts by identifier, rollback does not require a code deployment, a review cycle or a hunt for the old text. The owner changes one value, and within minutes every new request uses the previous version. We decide the rollback triggers before release, not during an incident. Typical ones: the validation failure rate for the feature rises above its normal band; the fallback rate rises, meaning users are seeing the degraded experience more often; a critical check fails on a production input that the evaluation set did not cover. That third case always ends the same way: roll back, add the input to the evaluation set, fix, re-test, release again. The set gets stronger with every incident. Prompts and model changes together Prompts and models are coupled. A prompt tuned on one model can behave differently on another, even a newer and supposedly better one. So we never change both at once. When a product evaluates a new model, it runs the current live prompt on the new model against the existing evaluation set. If the results are good, the routing rule changes and the prompt stays the same. If the results are close but not quite there, a new prompt version may be written for the new model, and that pair is evaluated together as a unit. Changing one thing at a time is slower on paper. In practice it is faster, because when something goes wrong you know which change caused it. A worked example: a rule change at KeşifAtlası KeşifAtlası helps people understand visa and relocation eligibility, starting from Türkiye. The eligibility decision comes from a verified rules database, never from a model. The model's job is to explain, in plain language, which rules applied and why, citing rule identifiers. See Rules engines for eligibility. Imagine a destination country changes the minimum income requirement for one of its visa categories. The rules team updates the rules database first, with the new threshold, an effective date and a source. That is a data change, not a prompt change. Now suppose the explanation prompt, rules.explain, contains an example that mentions the old threshold as a sample figure. That example is part of the prompt version. Left alone, it could nudge the model toward repeating the outdated number in some explanations. So the owner creates a new version with a neutral example, and the evaluation set gains new cases: an applicant just above the new threshold, one just below it, and one who would have qualified under the old rule but no longer does. The new version is scored. The critical checks: every figure in the explanation must come from the input, every rule cited must exist in the input's rule list, and no explanation may state or imply a decision different from the one the rules engine produced. The new version passes. It rolls out to a share of traffic, the numbers hold, and it goes live. The old version stays in the registry, unchanged, in case it is ever needed. The lesson generalises. Keep facts out of prompts wherever possible. When a fact must appear, it belongs to the version, and changing it is a release. What this looks like in a small team None of this requires a large platform team. The registry can start as a structured file per product with versions and changelogs, and a script that runs the evaluation set and prints a table. That is roughly where ours started. We built more only when the number of prompts made the simple version painful, in keeping with the rule against architecture theatre described in What an AI gateway does. What matters is the discipline, not the tooling: prompts out of application code; immutable versions with owners and changelogs; an evaluation set per feature, growing from real failures; a written release gate; a rollback that is one configuration change. A founder with one AI feature can adopt all five this week. Ownership, review and the changelog Every prompt in the registry has one owner. Not a team, not a channel: one person. Others can propose versions, but the owner approves what goes live and answers for how the feature behaves. For low-risk features, the owner's own review is enough once the evaluation passes. For features where a wrong output could mislead someone about something that matters to them — an eligibility explanation at KeşifAtlası, a learning explanation for a student at EduRelia, anything that touches consent or contact details at CastLyra — a second person reviews the diff and the evaluation table before release. The reviewer is looking for two things the numbers can miss: instructions that quietly widen what the model is allowed to say, and examples that smuggle in facts. The changelog is where that review leaves a trace. A good entry is short and specific: Version 12. Removed the sample income figure from the second example; added three threshold cases to the evaluation set. Critical checks unchanged; clarity improved on edge cases. Reviewed by the product lead. A poor entry says "tweaks" or "improved tone". Six months later, nobody can tell what changed or why, and the version history stops being useful. We also keep a short note on each prompt describing what it must never do. For a briefing prompt: never state a figure that is not in the input. For an explanation prompt: never state or imply an eligibility outcome. These lines turn directly into critical checks in the evaluation set, so the boundary is tested every release rather than remembered. Common mistakes Editing live prompts in place. It destroys the history you need for rollback and debugging. Evaluating by reading. A prompt that looks better can behave worse. Run the set. Only testing the happy path. Most production incidents come from edge and adversarial cases. Using real customer data in test sets. Anonymise or synthesise. Test data is still data, and it has to follow the same retention and deletion rules as everything else. Letting a model grade itself. Graders need their own validation against human judgement. Changing the prompt and the model together. You lose the ability to tell which one caused a change. Shared prompts across products. It feels efficient and quietly couples products that should stand alone. Each Oryvelon company owns its prompts, consistent with shared infrastructure, separate data. How it pays off The benefits compound over time. Every incident adds a test case, so the same failure rarely happens twice. Every model release can be evaluated quickly, because the test sets already exist. Every change has an author and a reason, so a new team member can read the changelog and understand why a prompt looks the way it does. And the cost side improves too: a clean evaluation set is what lets a team prove a smaller, cheaper model is good enough, which is one of the most reliable levers in cost discipline for AI products. It also changes the conversation inside a team. "I think the new prompt is better" becomes "version 9 matches version 8 on every critical check and improves clarity scores on the edge cases". Disagreements get shorter. Summary Prompt versioning and evaluation treat production prompts as what they are: code that changes behaviour for every user at once. At Oryvelon each product keeps its own prompt registry in the AI gateway, with immutable numbered versions, owners, changelogs and test status. Every feature has an evaluation set built from typical, edge, known-failure and adversarial cases, scored with mechanical checks wherever possible. New versions pass a written release gate, roll out to a share of traffic first, and can be rolled back with a single configuration change. The practice needs little tooling, grows stronger with every incident and turns prompt changes from opinion into evidence. ## Structured outputs: validating AI responses before users see them URL: https://oryvelon.com/insights/structured-outputs-schema-validation/ Category: AI Operations Free-text AI output is fine when a person reads it. It becomes a problem the moment code has to act on it. A dashboard needs a list of three priorities, not a paragraph that sometimes contains three priorities. A workflow needs a category from a fixed set, not a creative synonym. A database needs a valid identifier, not a product name the model half-remembers. Structured outputs solve this by giving each AI response a contract: a JSON schema the output must match. Schema validation enforces the contract before anything reaches the application. At Oryvelon, validation is one of the standard jobs of our shared AI gateway: application code in every company receives either a valid object or a clear failure, never something in between. This article explains how we design schemas, what we check beyond the schema, and how retries and fallbacks fit together. Why free text breaks products Teams often start by asking a model for "a JSON object with these fields" in the prompt and parsing whatever comes back. It works in the demo. In production, the failure modes pile up: the model wraps the JSON in an explanation, or in a code block; a required field is missing on some inputs; a number arrives as a string, or with a currency symbol; a category comes back as a near-synonym of an allowed value; an array has five items when the screen has room for three; a field refers to a product, rule or record that does not exist. Each of these is individually easy to handle. Together they spread defensive parsing code across the application, and every screen ends up with its own slightly different idea of what a valid response looks like. Worse, the last failure — an invented reference — often parses perfectly. It is valid JSON pointing at something that is not real. Validation in one place fixes the first five. Reference checks fix the sixth. Designing a schema the model can actually meet Most model providers now support some form of constrained or schema-guided output, which makes format errors much rarer. It does not make schema design less important. A badly designed schema still produces badly designed data, just in valid JSON. Our rules for a feature's schema: Keep it small. Ask for the fields the application needs and nothing else. Every extra field is one more thing the model can get wrong and one more thing to test. Close it. Disallow additional properties. If the model adds a field nobody asked for, that is a sign it is doing something unexpected, and the response should fail. Prefer enums to free text. If a message can be one of six categories, the schema lists the six. "Refund request" and "refund_request" and "Refunds" are three different strings to code. Use identifiers, not names. When the output refers to something in the input — a product, an order, a rule — it cites the identifier the input supplied. Names are ambiguous and easy to paraphrase. Identifiers can be checked. Bound everything. Arrays have a maximum length. Strings have a maximum length. Numbers have ranges where ranges make sense. Mark what is required. Optional fields invite the model to skip the hard ones. Separate data from prose. If a response needs both a figure and a sentence about it, the figure is its own field, copied from the input, and the sentence is another. Code uses the figure. Users read the sentence. Here is the shape of a briefing priority in plain terms: an object with a signal_id that must match one of the signals in the input, a severity from a three-value enum, a headline of at most 90 characters, and an explanation of at most 300. No other properties allowed. At most three priorities per briefing. Validation is more than a schema check A schema tells you the output has the right shape. It cannot tell you whether the contents are true. We run validation in three layers. Layer 1: structure. Does the response parse, and does it match the schema — types, required fields, enums, lengths, no extra properties? Layer 2: references. Does every identifier in the output exist in the input? If the model cites signal_id: s_14 and the input only contained signals s_1 to s_9, the response is rejected. This is the single most effective check against hallucination we have, because it turns a vague problem ("the model made something up") into a precise, mechanical one ("this ID was not supplied"). Layer 3: business rules. Feature-specific checks that the schema cannot express. Any figure quoted in a text field must match a figure in the input. A severity of "high" requires a signal flagged as high by the product's own logic. An explanation for KeşifAtlası must not state an eligibility outcome different from the one the rules engine produced. Layer 3 checks are written by the product team, because only they know what "correct" means for their product. The gateway just runs them in a consistent place and records the results. Rejecting hallucinated references in practice The reference check deserves a closer look, because it shapes how we design the whole feature. The pattern is simple. Before calling the model, the application assembles the facts the model is allowed to talk about, each with an identifier. After the call, every identifier in the output is checked against that set. Anything outside it fails. For this to work, three design choices have to be made up front: The model never looks things up. It receives facts; it does not fetch them. If the model can query a catalogue or a database during generation, the set of valid identifiers is no longer known in advance. The input is complete for the task. If a briefing should be able to mention any product with low stock, then every low-stock product must be in the input. A model given a partial list will sometimes try to fill the gap. Identifiers are opaque. An ID like p_2931 is harder for a model to invent plausibly than a name like "Blue Canvas Tote", and far easier to check. This is the practical mechanism behind our principle that AI explains, verified data decides. The verified systems decide which facts exist. The model can only point at them. Retrieval-based features use the same idea: retrieved passages carry identifiers, and the output must cite passages that were actually retrieved for this request and this product. See Retrieval without leakage. Retries: bounded and deliberate When a response fails validation, the obvious move is to try again. Retries are useful, but unbounded retries are one of the most common ways AI features become slow and expensive without anyone noticing. Our defaults: One retry for most features. Validation failures on a well-designed prompt are usually random, and a second attempt often succeeds. The retry may include the error. For structural failures, telling the model which field was missing can help. For reference failures, we usually retry without extra hints, because "you cited an ID that does not exist" sometimes leads the model to swap in a different wrong one. Retries count against the feature's budget and latency limit. If the first attempt already used most of the time allowance, the retry is skipped and the fallback runs. Every retry is logged with its reason, so a rising retry rate is visible on the product's weekly health check. A feature where retries are frequent has a design problem, not a retry problem. It usually means the schema is too complicated, the prompt is ambiguous, the input is incomplete or the model tier is too small. The fix belongs in the prompt registry or the routing rule, not in more retries. See Prompt versioning and evaluation for production AI and Model routing: matching tasks to the right AI model. What the application does when validation fails After the retry, if the response still fails, the gateway returns a clear failure signal. The application then follows the feature's fallback, decided in advance. The fallback depends on the feature: Feature type Sensible fallback Classification Rule-based tagging, or mark as "unclassified" for a person to review Extraction Leave fields empty and ask the user to fill them in Summary over verified data Show the underlying data in a plain table without narrative Explanation of a decision Show the decision and the cited rules without the prose explanation Generated copy Show nothing new; keep the existing copy What every fallback has in common is that the invalid output is never shown. A half-valid response is more dangerous than no response, because it looks authoritative. We describe the broader approach in Fallbacks and degraded modes for AI features. A worked example: a refund spike at MerchNivo MerchNivo is the AI e-commerce employee for Shopify stores. Imagine a store where one product suddenly receives several refund requests in a single day. MerchNivo's own code detects the spike from the store's data. It creates a signal with an identifier, the product's ID, the refund count, the usual daily range and the refund reasons customers gave. That signal goes into the input for the daily briefing, alongside the other signals for the day. See Operations signals for Shopify stores. The model returns a briefing. Validation runs: Structure. Three priorities, each with the required fields, severities from the enum, lengths within limits. Pass. References. The first priority cites the refund signal. The second cites a low-stock signal. The third cites a product ID that was not in the input — the model has linked the refund spike to a "similar product" it seems to have invented. Fail. The gateway retries once without hints. The second response cites only supplied identifiers. Layer 3 then checks that the refund count mentioned in the explanation text matches the count in the signal. It does. The briefing is delivered. If the retry had also failed, the merchant would have seen the same signals in a plain list — the refund spike, the low-stock warning — without the narrative. The facts would still arrive. Only the prose would be missing. The merchant never sees the invented product. That is the whole point. Structured outputs across our companies The same mechanism serves very different products. KeşifAtlası requires every explanation to cite rule identifiers from the verified rules database, and a Layer 3 check confirms the explanation does not contradict the rules engine's outcome. An explanation that sounds right but cites no rules is rejected. CastLyra uses AI to help structure talent profiles and brand briefs. Its schemas keep contact details out entirely — there is no field for them — and a Layer 3 check rejects generated profile text that mentions credentials, agencies or past clients not present in the verified profile data. See Verification in talent marketplaces. EduRelia uses schemas to keep learning content within the lesson's scope: a question object references a curriculum item from the input, answer options come from a bounded list, and explanations are length-limited and checked for age-appropriate language. ZodiVela uses structure mostly for safety and consistency. A reading has defined sections with length limits, and post-response checks keep the language within its framing of reflection and entertainment rather than prediction or advice. See Positioning consumer AI as entertainment. Monitoring validation over time Validation produces some of the most useful operational data an AI product has. Every failure is labelled: which layer, which field, which reference, which prompt version, which model. We watch three things: Validation failure rate per feature. A sudden rise usually means something changed: a prompt version, an input format, a model update on the provider side. Failure mix. Structure failures point at schema or prompt problems. Reference failures point at incomplete inputs or a model tier that is too small. Business-rule failures point at prompt instructions that need tightening. Fallback rate. How often users actually see the degraded experience. That is the number that matters to them. Failures with personal information removed become cases in the feature's evaluation set, so the next prompt version is tested against exactly the inputs that tripped up the last one. Changing a schema without breaking the product Schemas change. A screen gains a field, a category is split in two, a limit turns out to be too tight. Because the schema is the contract between the model and the application, a careless change breaks both sides at once. We version schemas alongside prompts. A prompt version in the registry names the schema version it was written for, and the application declares which schema versions it can read. A change then follows a predictable order: Additive changes first. A new optional field is added to the schema, and the application learns to read it while still working without it. The prompt follows. A new prompt version starts producing the field, and it is evaluated against the feature's test set like any other release. Tighten last. Only once the new field is reliably present does it become required, in a further schema version. Removing or renaming a field runs the same sequence in reverse. Splitting an enum value — say, "delivery issue" becoming "late delivery" and "damaged in transit" — is treated as a breaking change, because every downstream report that counted the old value needs to know. The rule underneath is the same as for prompts: a published schema version is never edited. Logs record which schema version each response was validated against, so an old failure can always be understood in its own terms. A checklist for a new structured feature Before an AI feature that returns data goes live in one of our companies, the team can answer these: What exactly does the application do with each field? If nothing, remove it. Is every category an enum, every array bounded, every string limited? Which fields refer to input items, and do they use identifiers? Does the input contain every item the model might legitimately need to cite? What are the business-rule checks, and who wrote them? How many retries, and what is the latency limit including them? What does the user see if validation fails twice? Are failures logged with the layer and field, without storing raw personal data? If any answer is unclear, the feature is not ready. Common mistakes Trusting "valid JSON" as "valid output". Shape is not truth. Check references and business rules. Open schemas. Allowing extra properties lets unexpected behaviour through silently. Free-text categories. Every synonym becomes a bug somewhere downstream. Letting the model look things up. It makes the set of valid references unknowable in advance. Retrying until it works. It hides design problems and inflates cost and latency. Showing partially valid output. If it failed, it does not ship. Validation spread across the application. One place, one set of rules, consistent logs. Summary Structured outputs give every AI response a contract, and validation enforces it before anything reaches users. We keep schemas small, closed and built on enums and identifiers; check structure, references and business rules in three layers; reject any output that cites something not supplied in the input; allow one deliberate retry; and fall back to a non-AI view of the verified data when validation still fails. Running this in a shared gateway gives every Oryvelon company consistent protection against malformed and hallucinated responses, and turns every failure into a test case that makes the next version better. ## Earning user trust in AI products during onboarding URL: https://oryvelon.com/insights/ai-product-onboarding-trust/ Category: AI Operations Users do not arrive at an AI product neutral. Some expect magic and will be disappointed by anything less. Others expect it to be wrong and are looking for the first mistake. Most carry a bit of both. Whatever they expect, they make up their minds quickly, and they do it based on what the product shows them, not what the marketing promised. Building trust in an AI product during onboarding is therefore a design problem with a few concrete parts: setting honest expectations, showing where answers come from, asking for as little as possible, and reaching a first piece of real value fast. Across Oryvelon's companies the products are very different — a store assistant, an eligibility checker, a learning tool, a reading app, a talent marketplace — but the onboarding principles are shared. This article sets them out, with examples. Why the first session decides so much AI products carry a trust burden that ordinary software does not. A spreadsheet does not make things up. A booking form does not sound confident about something it has wrong. When an AI product gets something wrong early, users do not file it as a bug. They conclude that the product cannot be relied on, and they stop checking whether later answers are better. The reverse is also true. A user who sees the product be accurate, modest about its limits and clear about its sources in the first few minutes tends to extend some patience when it later makes a mistake. They have a model of how it works. They know where to look to check it. So onboarding is not a welcome tour. It is the period in which the user builds their mental model of what the AI is, and the product has a short window to make that model accurate. Principle 1: say what the AI does, and what it does not The most useful thing an onboarding flow can do is tell the user, early and plainly, what the AI is responsible for and what it is not. We write this as a short statement for every AI product, in the product's own voice, and use it in onboarding, in help pages and next to AI-generated content. The statements come straight from each company's source-of-truth document, which already defines what the company will and will not do. Some examples, in paraphrase: MerchNivo: "Your numbers come from your store. MerchNivo's assistant explains them and suggests what to look at first. It does not change anything in your store without your approval." KeşifAtlası: "Eligibility results come from our verified rules database. The assistant explains which rules apply to you. It does not make the decision, and difficult cases are reviewed by a person." ZodiVela: "Readings are for reflection and entertainment. They are not predictions and not advice." EduRelia: "The assistant helps you practise and understand. Your teacher and school remain in charge of your learning." Notice what these avoid: capability boasting, vague phrases like "powered by advanced AI", and anything the product cannot keep. A limit stated upfront reads as honesty. The same limit discovered by the user reads as a failure. Principle 2: separate what is generated from what is verified A trustworthy AI screen makes it obvious which parts were written by a model and which came from verified data or a person. In practice this means consistent visual treatment. Figures pulled from a store, rules pulled from a database, profile details confirmed through verification — these are shown as data, in tables, labels or cards. AI-generated explanation is shown as explanation, visually distinct and labelled as such. Users quickly learn to read the two differently, which is exactly what we want them to do. This mirrors how the products are built. Under our principle that AI explains, verified data decides, the facts never originate in the model. The interface just makes that architecture visible to the user. Principle 3: show sources from the first screen If an answer rests on something, the user should be able to see what. For MerchNivo, a briefing line about refunds links to the orders it summarises. For KeşifAtlası, an explanation cites the specific rules, with their effective dates, and the user can open each one. For EduRelia, an explanation points to the lesson or curriculum item it relates to. Showing sources in onboarding matters more than showing them later, because it sets a habit. A user who clicks through to a source in their first session learns that the product can be checked. That changes how they treat every later answer: less blind belief, less blanket suspicion, more informed use. It also keeps the product honest internally. If a screen cannot show a source for a claim, that is a sign the claim should not be there. The mechanism that makes this reliable — outputs that must cite identifiers supplied in the input — is described in Structured outputs: validating AI responses before users see them. Principle 4: ask only for what the first task needs Onboarding flows love to collect data. Every extra field feels harmless on its own. Together they make the product feel greedy, and in AI products users are right to wonder what the data will be used for. Our rule is that onboarding asks only for what the first useful task requires. Everything else is requested later, in context, when the user can see why it is needed. For a MerchNivo store, the first briefing needs read access to orders, products and inventory. It does not need write access to anything. Write permissions for suggested actions are requested later, feature by feature, when the merchant chooses to use them. For KeşifAtlası, the free eligibility test asks only the questions the rules need to produce a result. Documents and detailed history belong to the paid report, if the user chooses it. See From free test to paid report. For EduRelia, a student profile holds the minimum the learning experience requires, and schools decide what is set up. See Data minimisation for children in edtech. Consent works the same way. Product consent — what the product needs to function — is asked separately from marketing consent, and neither is bundled into the other. We explain the difference in Marketing consent vs product consent. Principle 5: get to first value quickly — and make it checkable The strongest trust signal is a useful result the user can verify. Not a demo on sample data that could be staged, and not a promise of what the product will do once fully set up. We design each product's onboarding backwards from its first value moment: Company First value moment Why it builds trust MerchNivo A first briefing on the store's own recent activity The merchant knows their store and can check every figure KeşifAtlası A free test result with the rules that produced it The user can read the cited rules themselves EduRelia A practice explanation tied to a topic the student is studying The student and teacher can see how it matches the lesson ZodiVela A first short reading, clearly framed Expectations are set before any content appears CastLyra A structured profile draft from the talent's own details Nothing is added that the talent did not provide The best first value moments rest on data the user already knows well. A merchant who sees yesterday's refund count correctly stated, with a sensible observation attached, has just done their own quality check on the product. If first value requires setup that takes time — a store sync, a school licence, identity verification on a marketplace — we say how long it will take and show progress honestly. A clear "your first briefing will be ready after the sync, usually within the hour" is better than a spinner. Consumer products need the same care in a different register. At ZodiVela, first value is a short reading, but the framing comes before the content: a single line explaining that readings are for reflection and entertainment, shown before the first one appears rather than tucked into terms. Credits and subscriptions are explained in plain numbers before anything is charged. People enjoy a product more when they know exactly what it is, and they are far less likely to feel misled later. See Positioning consumer AI as entertainment. Principle 6: be honest about uncertainty and limits in the moment Expectations set on the first screen fade. They need reinforcing at the points where they matter. When the AI is less sure, the product says so. When a question is outside what the product covers, it says that too, rather than attempting an answer. When something needs a person, the route to that person is visible. See Designing human escalation paths in AI products. We avoid confidence theatre in both directions. No percentage scores that imply a precision the system does not have. No excessive hedging on every line, which trains users to ignore warnings. A limit is stated once, clearly, where it applies. A worked example: a merchant's first day with MerchNivo Imagine a merchant installs MerchNivo on their Shopify store. Minute one. The install screen explains, in three short lines, what MerchNivo reads (orders, products, inventory), what it does not do (change anything without approval) and where its numbers come from (the store). Permissions requested are read-only. Minutes two to five. While the store syncs, the merchant sees a short example briefing, clearly labelled as an example, showing what a briefing looks like: figures in one style, explanation in another, each line linked to its source. It also shows what happens when there is not enough data, so the first real "not enough data yet" message is not a surprise. First real briefing. Once the sync completes, the merchant gets a briefing about their own store. Every figure is clickable. The explanation is short. One suggested action is offered, with a note on why and a clear "do nothing" option. The first mistake. Imagine the merchant disagrees with a suggestion — perhaps the assistant flags a product as slow-moving that the merchant knows is seasonal. The product offers a simple way to say so, and the note changes what the merchant sees next time. The assistant's reasoning was visible, so the disagreement is about judgement, not about whether the product is inventing things. Nothing in that sequence depends on persuasive copy. It depends on the product being accurate, modest and checkable from the first minute. The store's own numbers do most of the persuading. When the user is not the buyer Onboarding gets harder when several people need to trust the product for different reasons. Two of our companies show this clearly. At EduRelia, the school often chooses the product, the teacher introduces it and the student uses it, while families want to understand what it does with their child's learning. A single welcome screen cannot serve all four. So each gets its own short explanation, in its own terms. The school sees what data is processed, where limits sit and what it controls. Teachers see what the assistant will and will not do in a lesson and how to review what students see. Students get a simple, age-appropriate statement of what the helper is for. Families get a plain summary they can read without a login. The content differs; the facts underneath are identical, drawn from the same source-of-truth document. See B2B2C licensing for schools. At CastLyra, talent and brands arrive with opposite worries. Talent want to know that their contact details, images and availability will not be handed to anyone who asks. Brands want to know that profiles are real. Onboarding for talent therefore starts with consent and contact-detail access rules — who can see what, and when — before any AI helps draft a profile. Onboarding for brands starts with what verification means on the platform and what it does not guarantee. The AI profile assistant appears only after those foundations are clear, and it works only from details the talent has supplied. See Verification in talent marketplaces. The general lesson: map every party whose trust the product depends on, write down what each one needs to hear first, and make sure none of the messages contradict each other. A product that tells a school one thing and a family another will lose both. What we avoid in onboarding A few patterns erode trust in AI products quickly, and we keep them out deliberately. Anthropomorphic overselling. Giving the AI a personality that implies more understanding than it has. Friendly is fine; pretending to be a person is not. Hidden AI. Presenting generated text as if a person wrote it, or as if it were verified data. Dark-pattern consent. Pre-ticked boxes, bundled permissions, or making the product unusable without optional data sharing. Cross-product familiarity. An Oryvelon product never greets a user with knowledge from another Oryvelon product. There is no shared user data across companies, and onboarding must not imply otherwise. Staged demos presented as real. Example content is always labelled as an example. Tracking the user harder than the product needs. Onboarding analytics measure whether people reach first value, not everything they do. Measuring whether onboarding builds trust Trust is hard to measure directly, but its effects are visible. The signals we look at for each product's onboarding: Time to first value. How long from sign-up to the first real, useful result. Share reaching first value. How many new users get there at all. Source engagement. Whether users open sources in the first sessions. Some checking is healthy; none at all may mean sources are hidden, and heavy checking of the same item may mean something looks wrong. Early corrections and disagreements. What users push back on, and whether the product handles it well. Return in the first week. Whether people come back after the first session. We read these as part of each company's 30/60/90-day checkpoints, and they feed straight into product decisions. A product where users reach value quickly but never return has a different problem from one where they never reach value at all. A checklist for AI onboarding Is there a plain statement of what the AI does and does not do, visible before first use? Can users tell generated content from verified data at a glance? Does every factual claim show its source? Does onboarding ask only for what the first task needs? Is there a first value moment based on data the user can verify? Are waiting times stated honestly? Is the route to a person visible where it matters? Is example content clearly labelled? Are product and marketing consent separate? Summary Trust in an AI product is decided early, by what users see it do rather than what it claims. Our onboarding across Oryvelon's companies follows the same principles: state plainly what the AI does and does not do, visually separate generated explanation from verified data, show sources from the first screen, ask only for the data the first task needs, reach a checkable first value quickly, and be honest about uncertainty and the route to a person. The result is users who understand how the product works, check it when it matters and give it the benefit of the doubt when it earns it. ## Designing human escalation paths in AI products URL: https://oryvelon.com/insights/human-escalation-in-ai-products/ Category: AI Operations Every AI product reaches questions it should not answer on its own. Some are too consequential. Some fall outside the rules it was built on. Some involve a person who is upset, confused or at risk. The question is not whether those moments happen, but whether the product was designed for them. Human escalation in AI products is that design: deciding in advance when the AI stops and a person takes over, how the handover works, what the user is told, and how each case makes the system better. It is not an afterthought or a "contact us" link at the bottom of the page. At Oryvelon, escalation is written into each company's source of truth before the AI feature is built. This article describes how we approach it, with three companies that face very different kinds of risk: KeşifAtlası, CastLyra and EduRelia. Why escalation belongs in the design, not the support queue Teams often treat escalation as a support problem. The AI handles what it can; anything else ends up in a general inbox. That arrangement fails in predictable ways. The AI does not know it should stop, so it attempts answers to questions it should have passed on. When a case does reach a person, it arrives without context, and the user has to explain everything again. Nobody has decided who is responsible for which kind of case, so urgent ones wait behind routine ones. And because escalations are not tracked as a category, the product never learns from them. Designed escalation fixes each of those. It gives the AI explicit reasons to stop. It hands the person a structured case. It assigns ownership. And it treats every escalation as information about where the product, the rules or the prompts need work. It also follows directly from how we build AI products. If AI explains, verified data decides, then there must be a defined answer to "what happens when verified data does not cover this case?" Escalation is that answer. What triggers an escalation We define triggers per product, in writing, and review them at each 30/60/90-day checkpoint. They fall into five groups. 1. Consequential outcomes. Some decisions matter too much to leave to an automated explanation alone: a result that could change whether someone applies for a visa, a verification decision that affects someone's ability to work, a concern about a child. For these, a person reviews before or alongside the automated result, depending on the product. 2. Gaps in the rules or data. The verified rules database has no rule for this combination of circumstances. The store data is incomplete. The curriculum does not cover the question. The AI must not fill a gap with a plausible guess. 3. Disputes. The user disagrees with an outcome, or reports that something the product showed them is wrong. Disputes always go to a person. 4. Explicit requests. The user asks for a person. We honour that without making them argue with the AI first. 5. Safety and wellbeing signals. Language suggesting distress, a risk to someone, a child in a situation that needs an adult, harassment on a marketplace. These go to people quickly, with the product's safety procedures, and the AI's role becomes limited to acknowledging and routing. See Safety boundaries for consumer AI products. Why model confidence is not enough It is tempting to trigger escalation mainly on model confidence: if the AI is unsure, hand over. Confidence signals are useful, but they are not sufficient. Models can be fluent and wrong, and a confident answer to a question the product should never answer is worse than an uncertain one. So most of our triggers are about the situation, not the model. A disputed eligibility result goes to a person whether the AI sounded confident or not. A question about a rule that is not in the database goes to a person even if the model could write a convincing paragraph. Confidence is one input among several, and it is never allowed to override a situational trigger. How a good handover works A handover has three audiences: the user, the person receiving the case, and the system that will learn from it. For the user, the product says clearly that a person will look at this, what they can expect and roughly when. It does not pretend the handover is instant if it is not. It does not ask the user to repeat what they have already said. For the reviewer, the case arrives structured: the user's question or the situation, in their own words where appropriate; the verified facts the product already has for this case; the rules or data that applied, and where the gap or conflict is; what the AI said, if anything, clearly marked as AI output; the trigger that caused the escalation; a priority, based on the trigger type. The summary is built largely from structured data rather than generated prose, for the same reasons we validate AI outputs everywhere else. See Structured outputs: validating AI responses before users see them. Where the AI does write a short summary for the reviewer, it is labelled as such, and the underlying facts sit beside it. For the system, every escalation is logged with its trigger, its outcome and what, if anything, should change: a new rule, a prompt adjustment, a new test case, a clearer onboarding message. Escalation without data leakage A handover moves information to a person, so it has to follow the same data rules as the rest of the product. The reviewer sees what the case needs and nothing more. Access runs through a defined role with least-privilege access and two-factor authentication. The case stays inside the product where it started: a KeşifAtlası case is handled in KeşifAtlası's systems, by people working for KeşifAtlası, and never enters a group-wide queue alongside other companies' users. There is no cross-product user data, and escalation is not an exception to that. See Shared infrastructure, separate data. Retention follows the product's rules. Escalated cases can contain sensitive details, so they have defined retention periods and deletion processes. KeşifAtlası: when the rules run out KeşifAtlası helps people understand visa and relocation eligibility, starting from Türkiye. Eligibility results come from a verified rules database. The AI explains which rules applied and why. It never decides. Its escalation design follows from that structure. Rule gaps. Imagine a user whose circumstances combine two situations the rules database does not address together — say, a change of employment status partway through a qualifying period. The rules engine returns "not covered" rather than a result. The AI does not attempt an answer. The user sees a clear message that their case needs a person, and the case goes to a reviewer with the answers they gave, the rules that were evaluated and the specific point where coverage ran out. Disputes. A user believes their result is wrong. Perhaps they have read something on an official site that seems to contradict it. The dispute goes to a person who checks the rule, its source and its effective date. If the rule was outdated, the rules team updates the database and the case becomes a test case for the explanation prompt. See Prompt versioning and evaluation for production AI. Paid reports. For the paid report, a person is part of the process for the harder questions by design. The AI prepares the explanation; people review the cases the product marks as complex. That is a large part of what the user is paying for, and the product says so. Throughout, the framing matters. KeşifAtlası explains rules and routes hard questions to people. It does not present itself as legal advice, and escalation messages are written to keep that line clear. See Rules engines for eligibility. CastLyra: when trust between people is at stake CastLyra is a marketplace connecting verified creative talent — models, creators, creative professionals — with brands. AI helps structure profiles and briefs. The consequential moments are about people, not rules. Verification decisions. Checks can be partly automated, but a decision that would reject or limit a talent's profile is reviewed by a person. Automated checks can flag; they do not have the final word. See Verification in talent marketplaces. Consent and contact details. CastLyra has strict rules about when a brand can see a talent's contact details. If a brand's message tries to get around those rules — asking a talent to share a phone number before the platform's conditions are met, for example — the case is flagged for a person, not handled by an automated warning alone. Safety concerns. Reports of harassment, misleading briefs or pressure to work outside agreed terms go straight to people with the authority to act on accounts. The AI's role is limited to acknowledging the report and routing it. It does not assess the credibility of the person reporting. Profile disputes. If talent disagrees with how the AI summarised their profile, they edit it directly, and the edit wins. If a brand disputes a profile's accuracy, a person reviews it against the verified data. On a marketplace, trust is the product. A slow or opaque escalation path on a safety report does more damage than any number of good matches can repair. EduRelia: when a child is involved EduRelia provides AI learning support for students, families and schools, with strict child data minimisation. Escalation here has a particular shape, because the right "person" is often not a support agent but a responsible adult already in the child's life. Learning questions beyond scope. If a student asks about something outside the lesson's scope, or the assistant cannot give a reliable explanation, it says so and points the student to their teacher. It does not improvise. Wellbeing signals. If a student's messages suggest distress or a situation that needs an adult, the assistant stops being a tutor. It responds with care, directs the student to a trusted adult, and the school's agreed procedure applies. What happens next is defined in the school's licence agreement, not decided by the product on the fly. Content concerns. A teacher or parent who believes the assistant said something inappropriate can report it. The report goes to a person at EduRelia, the relevant output is reviewed, and any fix goes through the evaluation process before release. Minimal handover data. Because of the product's data minimisation rules, handovers carry as little as possible: the relevant exchange and the minimum context, not a profile of the student. See Data minimisation for children in edtech. What the user sees Across all three products, the escalation message follows the same pattern: What is happening. "A person will review this." Why. "Your situation isn't covered by the rules we can check automatically." Or: "You asked to speak to someone." Short and honest. What to expect. A realistic time frame and how they will be contacted, within the product. What they can do meanwhile. Where relevant, what information would help, or where to find official sources. We avoid two things in particular: pretending a person is involved when they are not, and making escalation feel like a failure on the user's part. Being passed to a person should feel like the product working as intended — because it is. Making the route to a person visible from the start is part of how we build trust during onboarding. See Earning user trust in AI products. Staffing escalation in a small company Escalation needs people, and our companies are designed to be lean. Three practices keep that manageable. Priorities by trigger. Safety signals and disputes on consequential outcomes are handled first. Rule gaps are handled in order. Explicit requests for a person get a prompt acknowledgement even when the answer takes longer. Escalation budgets. Each product tracks its escalation rate. A rising rate is a signal to fix something upstream — a missing rule, an unclear onboarding message, a prompt that attempts too much — rather than simply adding reviewers. Degraded modes that include people. When an AI feature is unavailable, some products can route affected requests to people for a period instead of failing entirely. That choice is made per product, and the capacity limits are written down. Learning from every escalation An escalation is a small, expensive event. It is worth extracting everything from it. Each case is tagged with its trigger and outcome. Monthly, each company's owner looks at the patterns as part of the regular monthly review: Rule gaps that recur become new rules in the database. Disputes that turn out to be right become fixes and test cases. Questions the AI attempted but should not have become new triggers. Confusion that stems from onboarding becomes clearer copy. Over time, the aim is not zero escalations. It is fewer avoidable ones, and fast, well-handled ones for everything that genuinely needs a person. Writing the escalation section before launch Every Oryvelon company's source of truth has a short escalation section, written before the first AI feature ships. It answers five questions: Which situations always go to a person? Listed by trigger type, with examples specific to the product. Who receives each type of case? A named role, not "the team". What does the reviewer see? The fields in the structured case, and anything explicitly excluded. What is the user told, and what response time is realistic? Written as the actual message copy, not a note to write it later. How does each case feed back? Which log, which review, which owner. Writing this down early changes the product. Teams discover triggers they had not considered, such as what happens when a user asks the same question three times in slightly different words, hoping for a different answer. They notice data the reviewer would need that the product does not currently keep, or data the product keeps that no reviewer should see. And they set response-time promises they can actually meet with the people they have, rather than ones that sound good on a landing page. The section is short, usually less than a page. It is reviewed at each 30/60/90-day checkpoint along with the escalation numbers, and it changes whenever the product's scope does. Common mistakes No written triggers. The AI keeps answering because nobody told it when to stop. Confidence-only escalation. Fluent, confident, wrong answers slip through. Context lost at handover. The user repeats everything, and the reviewer starts from zero. Hiding the route to a person. Users who cannot find it lose trust in everything else. Group-wide queues. Mixing cases from different companies breaks data boundaries. Escalation as a dead end. Cases are closed but never feed back into rules, prompts or tests. Summary Human escalation is the designed point where an AI product stops and a person takes over. We define triggers in writing for each product — consequential outcomes, gaps in rules or data, disputes, explicit requests and safety signals — and never let model confidence override them. Handovers carry a structured case with verified facts so users do not repeat themselves, stay inside the product that owns the data, and tell the user honestly what happens next. KeşifAtlası routes rule gaps and disputes to people, CastLyra keeps verification and safety decisions with people, and EduRelia hands wellbeing concerns to the responsible adults. Every escalation feeds back into rules, prompts and test sets, so the product gets better at knowing when to ask for help. ## AI fallbacks and degraded modes: what a product does when the model fails URL: https://oryvelon.com/insights/ai-fallbacks-and-degraded-modes/ Category: AI Operations Every AI feature fails sometimes. A provider has an outage. A request times out at the worst moment. A model returns output that does not match the schema, twice. A product reaches its monthly budget on the 26th. None of this is exotic; it is ordinary operations for anything that depends on a model you do not run yourself. The question is not whether a feature will fail but what the user sees when it does. At Oryvelon, every AI feature in every company ships with written AI fallbacks and a non-AI degraded mode. This note explains how we design them, why the degraded mode is usually more important than the fallback model, and what we learned about making failure boring. Why AI features fail differently Traditional software fails in fairly predictable ways. A database is up or down. An API returns an error code. You can usually tell what went wrong and when. AI features add three failure types that are harder to spot. Slow failure. The model responds, but after twenty seconds instead of three. Nothing is technically broken, and yet the user has left. Quiet failure. The model responds quickly with confident output that is wrong, incomplete or outside the allowed format. Without validation, this looks exactly like success. See Structured outputs and schema validation. Economic failure. The model works perfectly, but the product has spent its budget. Continuing to call it means losing money on every request. See Cost discipline for AI products. A fallback design that only handles hard errors covers maybe a third of the real cases. Ours treats all four — hard errors, slowness, invalid output and budget exhaustion — as triggers for the same planned behaviour. The principle that makes degraded modes possible Degraded modes are easy to design if you follow one rule from the start: the model is never the source of the facts. In our products, verified systems supply the facts and the model explains them. MerchNivo's numbers come from the Shopify store. KeşifAtlası's eligibility outcomes come from a verified rules database. CastLyra's verification status comes from its own review process. The model turns those facts into plain language, suggestions or a structured draft. We wrote about this in AI explains, verified data decides. The consequence for failure handling is enormous. If the model disappears, the facts are still there. The degraded mode can show them directly — as a table, a checklist, a status badge, a template — and the user still gets what they actually came for. They lose the narrative, not the truth. Products that let the model generate the facts have no such option. When the model fails, there is nothing to fall back to except an error message. Anatomy of a fallback chain For each AI feature, the product's configuration in our AI gateway defines a fallback chain. A typical chain has four steps. Step What happens Typical time budget 1. Limited retry Same model, same prompt, once. Handles transient errors and a single invalid output. Short, only if time remains 2. Approved alternative A second model or a simpler prompt version that has passed the same evaluation set. Short 3. Degraded mode The application renders the non-AI version of the feature from verified data. Near instant 4. Clear message If even the degraded mode cannot run, the user is told plainly what is unavailable and what to do. Instant Two details matter more than the steps themselves. First, the whole chain has a single deadline. If a feature should respond within a few seconds, the chain does not get a few seconds per step. It gets a few seconds in total, and each step only runs if enough time remains. Otherwise a chain of three careful fallbacks turns into a thirty-second wait, which is just a slow failure with extra steps. Second, the alternative model must be approved for that feature. It is tempting to say "if the main model fails, use any other one". We do not. A model that has not passed the feature's evaluation set can produce output that is worse in ways validation will not catch: a different tone, weaker adherence to content boundaries, subtle factual drift. For some features we do not configure an alternative model at all and go straight from retry to degraded mode. See Model routing and Prompt versioning and evaluation. Designing the degraded mode first We now design the non-AI version of a feature before, or alongside, the AI version. That sounds backwards. It turns out to be one of the most useful habits we have. Designing the degraded mode forces three questions early. What is the user's actual job here? If the degraded version still does the job, the AI is a quality layer on top. If it cannot, the feature may depend on the model in ways that are risky. Which facts does the feature need? Listing them makes it obvious which verified system provides each one, and exposes any fact that only the model was producing. What does "good enough" look like without AI? Often the answer is simpler and more useful than expected, and sometimes it becomes the default for certain users. A degraded mode is not an error state with a nicer icon. It is a real, maintained piece of the product with its own design, its own copy and its own tests. Worked examples across the group MerchNivo: the briefing becomes a table MerchNivo is an AI e-commerce employee for Shopify stores. Its best-known feature is the morning briefing: a short, plain-language summary of what happened in the store and what deserves attention. Imagine a store that had an unusual spike in refunds overnight — six refund requests, four of them for the same product. On a normal day, the briefing says something like: refunds are higher than usual, most relate to one product, you may want to check the listing and recent reviews. On a day when the model is unavailable, the merchant opens MerchNivo and sees the degraded briefing. The same signals appear as a compact table: refunds (6, above the recent range), top refunded product (4 of 6), low-stock items, abandoned checkouts. The operations signal that flagged the refund spike is a rule, not a model output, so it still fires. A small note says the written summary is temporarily unavailable. The merchant loses the prose and the suggested wording of a next step. They do not lose the fact that something is wrong with one product. That is the difference between a degraded mode and an outage. KeşifAtlası: the rules still answer KeşifAtlası helps people understand visa and relocation eligibility, grounded in a verified rules database. The AI explains the rules; it never decides. If the explanation model fails during a free eligibility test, the rules engine has already produced the outcome and the list of rules that applied. The degraded mode shows that outcome with the rule texts themselves and their official wording, plus a clear route to ask a person. It reads more formally, but it is complete and correct. For the paid report the bar is higher. We do not want a customer who paid for a clear, explained report to receive a rule list. So if generation fails for a paid report, the report is queued, the customer is told it will arrive shortly, and the job is retried in the background. If it still fails, a person reviews it. Degraded mode for a paid deliverable is often "later, and checked", not "now, but plainer". See From free test to paid report. ZodiVela: fewer words, same boundaries ZodiVela offers personal discovery readings framed for reflection and entertainment. Its degraded mode uses pre-written, reviewed content blocks for the relevant sign and theme, assembled by simple rules. They are less personal than a generated reading but fully within the product's content boundaries, because every block was written and reviewed in advance. One rule here is firm: if a user paid credits for a personalised reading and received the degraded version, the credits are not consumed. Charging full price for a fallback would turn a technical failure into a trust problem. EduRelia: stop, don't improvise EduRelia serves students, families and schools. For a learning product used by children, a wrong or inappropriate answer is worse than no answer. Its fallback chain is deliberately short: one retry, then the degraded mode, which points the student to the relevant lesson material and suggests asking a teacher or parent. There is no approved alternative model for the tutoring features unless it has passed the same age-appropriate content evaluation. See Safety boundaries for consumer AI products. Budget limits are a fallback trigger too Every product has a monthly AI budget with a soft limit and a hard limit. The soft limit sends an alert to the company owner. The hard limit switches the product, or specific features, into degraded mode. The alternative approaches are both worse. Letting spending continue past the limit turns an operations issue into a financial one. Returning errors when the budget is hit punishes users for something they cannot see or influence. In practice, hitting a hard limit rarely means switching everything off. Budgets are usually set per feature, so the least essential features degrade first. A product might keep its core explanation feature running while pausing optional extras such as alternative phrasings or long-form expansions. Heavy users who have hit a per-user allowance may see the degraded mode while everyone else continues normally. Each of those choices is written down in the product's configuration, not decided under pressure at 11pm. When a product hits its hard limit, it is also a signal for the monthly review. Was the budget wrong, or did usage change, or is a feature more expensive than it is worth? That feeds the continue, stop or scale checkpoints we run for every company. What the user should see The copy for a degraded mode deserves as much care as the copy for the main feature. A few rules we follow: Say what is missing, not what broke. "The written summary is not available right now" is more useful than "AI service error". Show what still works. Lead with the facts the user can rely on. Do not overpromise. Avoid "back in 5 minutes" unless you know it. "We'll try again automatically" is fine if true. Do not charge for what was not delivered. Credits, usage allowances and paid reports should reflect what the user actually received. Offer a human route where the stakes are high. For KeşifAtlası and EduRelia, the degraded mode always includes a way to reach a person. See Human escalation in AI products. Users are remarkably tolerant of a plainer experience when it is honest. They are much less tolerant of spinners that never finish. Observability: know when you are degraded A fallback that works silently has a hidden cost: nobody notices that the product has been in degraded mode for three days. The gateway records every step of the chain per product and per feature — how many requests succeeded on first try, how many used the retry, how many used the alternative, how many ended in the degraded mode. We watch the fallback rate as a first-class health metric. A small, steady rate is normal. A sudden jump usually means a provider issue. A slow climb usually means something on our side has drifted: a prompt change that increases invalid outputs, an input format that changed, a model update that altered behaviour. We also log the reason for each fallback in categories — timeout, provider error, schema failure, safety block, budget limit — so that the weekly health review can tell the difference between "the provider had a bad afternoon" and "our new prompt fails validation one time in ten". Test the fallback, not just the feature The most common fallback bug is that the fallback itself does not work. It was written once, never exercised, and broke quietly when something else changed. So we test fallbacks deliberately: In automated tests, every AI feature has a test that forces each step of the chain — simulated timeout, simulated invalid output, simulated budget limit — and checks what the user sees. In staging, we can switch a feature to degraded mode with a configuration flag and click through it like a user would. Before launch, the company owner looks at the degraded mode and answers one question: would I be comfortable if a paying customer saw only this for a day? If the honest answer is no, the degraded mode is not finished. Common mistakes A few patterns we have seen, in our own early work and elsewhere: Falling back to a bigger model. When a small model fails, reaching for a larger one feels safe. It is often slower, far more expensive and may not have been evaluated for the task. It can also hide a prompt problem that should be fixed. Unlimited retries. Retries multiply cost and latency. A single retry catches most transient issues; after that, move on. Caching a fallback as if it were the real answer. If a degraded response is stored in a cache under the same key as a full response, users can keep seeing the degraded version long after the model has recovered. Degraded output should be marked and cached separately, if at all. See Latency and caching for AI. One global switch. A single "AI off" switch for a whole product is useful in an emergency, but it should not be the only control. Per-feature fallbacks let the least important features degrade while the core keeps working. Hiding the degraded state from the team. If only the gateway knows, nobody fixes the cause. A fallback checklist for new AI features Before any AI feature ships in an Oryvelon company, its entry in the company's source-of-truth document answers these questions: What is the user's job, and does the degraded mode still do it? Which verified systems provide the facts, and do any facts come only from the model? What is the total time budget for the feature, including fallbacks? Is there an approved alternative model or prompt, and which evaluation did it pass? What exactly does the user see in degraded mode, and has someone reviewed that copy? What happens to credits, allowances or payments if the degraded mode is used? When does budget exhaustion trigger degradation, and for which features first? How is the fallback rate monitored, and who is alerted? Has each step of the chain been forced in a test? Nine questions, usually answered in a page. It is one of the cheapest reliability investments we make. Summary AI fallbacks and degraded modes are what separate a product that occasionally feels plainer from one that occasionally breaks. The design becomes straightforward once the model is never the source of the facts: when it fails, verified data can still be shown directly. We give every feature a short, time-bounded fallback chain of retry, approved alternative and non-AI degraded mode; treat slowness, invalid output and budget limits as failures alongside hard errors; write honest copy that never charges for what was not delivered; and monitor and test the fallback as carefully as the feature itself. ## Latency and caching for AI features: keys, batching and streaming URL: https://oryvelon.com/insights/latency-and-caching-for-ai/ Category: AI Operations Speed is a feature. A good AI answer that takes fifteen seconds loses to a decent one that takes two, and both lose to a plain table that appears instantly. Cost follows the same shape: the fastest request is often the one you never had to send. This note covers how we approach latency and caching for AI across Oryvelon's companies: where we measure, what we cache and how we build cache keys, when we batch, and when streaming is worth it. The mechanics sit in our shared AI gateway; the decisions sit with each product. Where latency actually comes from When someone says "the AI is slow", the model is usually only part of the story. A typical AI feature spends time in several places: Gathering facts. Querying the store, the rules database or the product's own tables to build the input. Building the prompt. Loading the registered prompt, adding the facts, applying pre-request safety hooks. Waiting in line. Rate limits, provider queues, network. Time to first token. How long the model takes to start responding. This grows with input length and model size. Generation. How long it takes to produce the full output. This grows with output length. Validation and post-processing. Schema checks, post-response hooks, possibly a retry. Rendering. Getting the result on screen. Most teams optimise step 5 and ignore the rest. In our experience the biggest gains usually come from steps 1, 4 and 6: fetching facts in parallel, sending less input, and avoiding retries by writing prompts that pass validation the first time. Measure what the user feels We measure latency from the moment the user acts to the moment they see something useful. The gateway also records the model call on its own, but the end-to-end number is the one that matters. And we look at the slow end, not the average. An average of three seconds can hide a tenth of requests that take twelve. Those are the ones people remember and complain about. The gateway reports slow-end latency per product and per feature, and it is one of the indicators each company owner sees in their weekly product review. Every interactive feature also has a latency budget written into its configuration: the total time it may take, including any retries and fallbacks. When the budget runs out, the feature moves to its degraded mode rather than keeping the user waiting. See AI fallbacks and degraded modes. The cheapest speedups Before caching or streaming, there are a few changes that almost always help. Send less. Input length drives both cost and time to first token. A MerchNivo briefing does not need every order from yesterday; it needs the aggregated figures and the few outliers worth mentioning. Summarise in code, then ask the model to explain the summary. Ask for less. Output length drives generation time. A cap on output tokens, plus a prompt that asks for three priorities rather than "a thorough analysis", often halves latency with no loss of usefulness. Use a smaller model for small jobs. Classification, extraction and short rewrites rarely need a large model. See Model routing. Fetch facts in parallel. If a feature needs data from three sources, fetch them at the same time. This is ordinary engineering and it is often the biggest single win. Get it right the first time. Every retry after a validation failure roughly doubles latency for that request. Tightening a prompt so it passes validation reliably is a latency fix. See Structured outputs and schema validation. Caching: what is safe to cache Caching model responses is powerful and easy to get wrong. The failure that worries us most is not a stale answer. It is the wrong person's answer. So we are conservative about what is eligible. A request can be cached only if: the output depends only on the input sent, not on hidden state, time of day or randomness the product relies on; the feature is low-variance by design — the same input should produce essentially the same useful output; the product has explicitly enabled caching for that feature in its configuration. Good candidates in our companies include explaining a specific rule in KeşifAtlası's rules database, generating a standard description for a product category in Noveniq's catalogue, or classifying the type of a support message. Poor candidates include anything personal, anything that combines one user's data with a question, and anything creative where variety is part of the value. Building a safe cache key The cache key is where safety lives. Ours is built from these parts: Key component Why it is there Product identifier A cached answer from one company can never be served to another. Tenant or scope identifier (where relevant) For B2B or multi-tenant features, entries are never shared between customers. See Tenant isolation explained. Feature name The same input used in two features gives two different outputs. Prompt identifier and version When a prompt changes, old answers stop matching automatically. Model identifier A model change produces different output; it should not reuse the old cache. Input fingerprint A hash of the exact, normalised input, including the facts that were sent. Relevant parameters Output limits, language, schema version. Two parts deserve a closer look. The prompt version. Including it means cache invalidation for prompt changes is automatic. When a new version of rule.explain goes live, every request produces a new key, and the old entries simply age out. Nobody needs to remember to clear anything. This is one of the reasons we register prompts with versions in the first place; see Prompt versioning and evaluation. The input fingerprint. We normalise the input before hashing — consistent ordering of fields, trimmed whitespace, stable number formatting — so trivially different inputs do not miss the cache. But we fingerprint the facts as well as the question. If KeşifAtlası updates a rule's text, the input changes, the fingerprint changes, and the cached explanation of the old text is no longer returned. The cache follows the truth instead of lagging behind it. We also set an expiry on every entry, even when the key should make staleness impossible. Keys can be designed imperfectly, and an expiry is a cheap safety net. Where caching goes wrong A few mistakes we design against: Caching personal data. If the input contains anything about a specific person, the cached entry is effectively a copy of personal data with its own retention period. We avoid it, and where it is truly needed, the entry is scoped to that user, short-lived and covered by the product's retention rules. See Data retention and deletion. Caching degraded output. If a fallback response is stored under the same key as a full response, users can keep receiving the fallback after the model recovers. Degraded output is either not cached or cached under a separate, short-lived key. Caching failures. A validation failure or safety block should never be cached as an answer. Semantic caching without care. Matching "similar" questions to cached answers is appealing, but similarity is not sameness. For a rules product, two questions that look alike can have different correct answers because one detail differs. We only use exact-match caching for anything where correctness matters. Sharing a cache across products for efficiency. Even if two products happened to ask identical questions, we would not share the entry. The saving is tiny; the precedent is bad. Shared infrastructure, separate data — see Shared infrastructure, separate data. Precompute what nobody is waiting for The fastest request is one that already happened. A surprising amount of AI work in real products does not need to run while a person waits. MerchNivo's morning briefing is the clearest example. Merchants open it in the morning. The underlying numbers are final shortly after midnight in the store's time zone. So the briefing for each store is generated in the early hours, validated, and stored. When the merchant opens the dashboard, it loads instantly. If generation failed overnight, there is time to retry before anyone looks. Catalogue content for Noveniq — draft product descriptions, alt text, attribute extraction — is produced in batches when products are added, then reviewed by a person before publishing. Nothing about it needs to be interactive. Evaluations of new prompts and models run as scheduled jobs against stored test sets, never on live traffic. Precomputing has its own cost: you pay for outputs some users will never look at. For a daily briefing that most merchants open, that is fine. For a detailed report that only a few people request, generating on demand is cheaper. We decide per feature, based on how often the output is actually used. Batching Batching means grouping many independent requests and sending them together, typically through a provider's batch interface or simply a queue that processes jobs steadily instead of all at once. It helps in three ways. Batch interfaces are often cheaper per request. A queue smooths load, so a burst of work does not hit rate limits. And failures are easier to handle: a failed item goes back in the queue instead of producing an error on someone's screen. The trade-off is time. Batched work may take minutes or hours to complete. That is fine for the overnight briefings and catalogue drafts above. It is not fine for anything interactive. A simple rule we use: if a person is looking at a spinner, it is interactive. Otherwise, it is a candidate for the queue. Streaming: helpful, but not everywhere Streaming sends output to the user as it is generated instead of waiting for the full response. It makes long responses feel much faster, because the reader starts reading within a second or two. It fits well where the output is prose and a person reads it top to bottom. A ZodiVela reading is a good example: the first paragraph appears quickly, and the reader is already engaged while the rest arrives. It fits poorly in three situations. Structured outputs. If the product needs a validated JSON object — MerchNivo's list of priorities with references to figures, KeşifAtlası's explanation with rule identifiers — it cannot show half of it safely. It needs the complete object, validated against the schema, before anything reaches the user. Content that must pass post-response checks. For consumer products with content boundaries, streaming means text reaches the screen before the whole response has been checked. We handle this for ZodiVela by running checks on each completed paragraph before it is displayed, with the stream held briefly while a paragraph finishes. For EduRelia, where users are students, we generally do not stream tutoring responses at all; the full answer is checked before it is shown. A slightly slower answer is a fair price. See Safety boundaries for consumer AI products. Short answers. If the whole output takes a second to generate, streaming adds complexity for no real benefit. A worked example: KeşifAtlası's rule explanations To show how these pieces combine, here is how a single feature is tuned. When a user completes KeşifAtlası's free eligibility test, the rules engine produces an outcome and a list of rules that applied. For each rule, the user can see a plain-language explanation. Facts first. The rules engine runs in code and returns instantly. The outcome is shown immediately, before any model call. Cache per rule. The explanation of a rule depends only on the rule's current text and the prompt version, not on the user. So explanations are cached with a key of product, feature, prompt version, model and a fingerprint of the rule text. Most requests are cache hits. Precompute on change. When a rule is added or updated in the verified database, its explanation is regenerated in the background, validated and reviewed. The first user to see it never waits. Personal part separate. Where the explanation needs to mention the user's own answers — "because you indicated…" — that part is assembled from templates in code rather than generated, so nothing personal ever enters the cache. No streaming. Explanations are short and structured, with rule identifiers the validator checks. The result is a feature that feels instant, costs very little per user, and updates itself when the rules change. See Rules engines for eligibility. Rate limits, queues and fairness Latency is not only about one request. When many requests arrive together, the product's own rate limits and the provider's limits decide who waits. Each product has limits per user and per feature, enforced by the gateway. That protects the budget, but it also protects latency for everyone else: one heavy user running dozens of requests in a minute should not slow the experience for the rest. On CastLyra, for example, a brand drafting several briefs in quick succession gets queued gracefully rather than blocking talent who are editing their own profiles at the same time. Interactive and background work also run in separate lanes. A large overnight batch for one product should never compete with a person waiting on a screen. Background jobs use their own queue with their own limits, and they back off when interactive traffic is high. Finally, provider limits are per product, because each company has its own AI project and keys. A traffic spike in one company does not consume another company's allowance, which is one more quiet benefit of keeping products separate. What we deliberately do not optimise Not every millisecond is worth chasing. We skip optimisations that add complexity without a clear effect on users or cost. We do not build custom model hosting to shave latency for small products. We do not add semantic caches, speculative generation or multi-model racing — sending the same request to two models and taking the fastest — unless a feature has a proven need. Racing, in particular, doubles cost to save a second, and it complicates evaluation because two models' outputs reach users. We also avoid tuning features nobody uses. The weekly metrics show which features carry real traffic. Those get attention. A feature used a handful of times a week can take a few extra seconds; if it turns out to matter, the numbers will say so. It is the same cost discipline we apply everywhere else: spend effort where it changes an outcome. A short checklist For each AI feature, we answer these in the company's source-of-truth document: What is the end-to-end latency budget, including fallbacks? Can any part of the work be precomputed or batched? Is the input as small as it can be? Is the output capped? Is caching enabled, and does the key include product, scope, feature, prompt version, model and input fingerprint? Does any cached content contain personal data? If yes, why, and for how long? Is streaming used, and if so, how are post-response checks applied? Which latency and cache hit indicators does the owner watch? Summary Latency and caching for AI are mostly about moving work to the right place. We measure the end-to-end delay users feel, especially at the slow end; shrink inputs and outputs and route small jobs to small models; precompute and batch anything nobody is waiting for; and cache only low-variance, non-personal requests using keys built from product, scope, feature, prompt version, model and an input fingerprint, so answers never cross users or products and refresh automatically when the prompt or the facts change. Streaming is used where people read long prose, not where outputs must be validated first. ## RAG without data leakage: per-product and per-tenant retrieval URL: https://oryvelon.com/insights/rag-without-data-leakage/ Category: AI Operations Retrieval-augmented generation — RAG — is the standard way to make a language model answer from your own material. You store documents in a searchable index, retrieve the most relevant passages for each question, and give them to the model as context. The model answers from what it was handed. That last sentence is also the problem. A model will use anything in its context. If retrieval hands it a document the user should not see, no system prompt can reliably stop that document from shaping, or appearing in, the answer. This note explains how we build RAG without data leakage across Oryvelon's companies: separate indexes per product, tenant separation where products serve organisations, access filtering before retrieval, and keeping indexes in step with the source of truth. It sits alongside our broader principle of shared infrastructure, separate data. Where RAG leaks It helps to be precise about how leakage happens, because most of it is not the model misbehaving. Retrieval returns the wrong documents. The index contains material from several users, tenants or products, and the query is not filtered. The most relevant passage belongs to someone else. The model uses it. This is the most common case by far. Metadata filters are optional. The filter exists, but a code path forgets to pass it — a new feature, an admin tool, a background job. The index happily returns everything. Filters come from the wrong place. The tenant or user identifier used for filtering is taken from something the user controls: the prompt, a request parameter, a field the model extracted. Anyone who can change it can read other people's material. The index outlives the source. A document is deleted or its permissions change in the product, but the copy in the index remains. Retrieval keeps serving it. Prompts are treated as access control. The system prompt says "only use documents belonging to this customer". The model mostly complies. Mostly is not a security property. Logs and caches hold the context. Even if retrieval was correct, full prompts with retrieved passages are logged or cached somewhere broader than the original documents. Every one of these is a design problem, and every one can be closed by design. Our rule: no group-wide knowledge The simplest decision we made was also the most important. There is no Oryvelon knowledge base. No shared index, no group-level document store, no "company brain" that products query. Each company has its own: document sources and ingestion jobs; index or indexes, in its own storage; embedding and model credentials, under its own AI project; retrieval configuration and prompts, registered under that product; logs and retention rules. The retrieval code is shared. It is a library that every product can use, and the AI gateway applies the usual routing, budgets and validation. But the library is always configured with one product's index and one product's credentials. It has no way to query across products because nothing it can reach spans them. Why so strict? Because the companies are genuinely different businesses with different users and different obligations. MerchNivo's knowledge about a merchant's store has no place in a ZodiVela reading. EduRelia's school materials have no place in CastLyra's brief drafting. And each company is designed so it could be separated from the group and keep running; see Designing every company to stand alone. An index shared across companies would make that separation a forensic exercise. Three kinds of knowledge, three levels of separation Within a single product, not all retrieved material is equally sensitive. We sort it into three kinds. Kind Examples Separation Product knowledge Help articles, verified rules, public lesson content, product policies One index per product, readable by all its users Tenant knowledge A school's own materials, a store's own policies and catalogue, a brand's briefs Separate index or namespace per tenant, or a mandatory tenant filter Personal knowledge A user's own notes, history, uploaded documents Scoped to the user within the tenant, often not indexed at all Most leaks come from mixing these kinds in one index with weak filtering. Keeping them apart structurally — different indexes or namespaces, not just a field on each record — removes whole categories of mistake. Separate indexes or a filtered shared index? For tenant knowledge, there are two common designs. Separate index or namespace per tenant. Each school, store or brand gets its own. Retrieval is pointed at the tenant's index by the application, based on the authenticated session. A query cannot return another tenant's documents because it never touches them. Deleting a tenant is deleting its index. One index with a mandatory tenant filter. All tenants' documents live together, each tagged with a tenant identifier. Every query must include a filter on that identifier. The second design can be efficient and is widely used. Its weakness is the word "must". If any code path forgets the filter, isolation fails silently. Where we use it, the filter is not a parameter the calling code adds; it is applied inside the retrieval library from the session context, and the library refuses to run a query without it. Our default is the first design where the number of tenants is manageable and deletion matters, and the second only with that enforced filter. See Tenant isolation explained for the broader pattern. Filters come from the session, never from the prompt This is the rule we repeat most often in code review. The identifiers used to scope retrieval — product, tenant, user, role — are taken from the authenticated session on the server. They are never taken from: the user's question or any text they typed; a request parameter the client can change; anything the model produced, such as an extracted "customer name". Consider a MerchNivo merchant who types: "Compare my return policy with the one used by the store at example-store.com." If the store identifier for retrieval came from the text, the merchant could read another store's policy documents. Because it comes from the session, the retrieval only ever looks inside the current store's own material. The model answers that it has no information about other stores — which is exactly right. The same logic applies to roles within a tenant. On EduRelia, a teacher and a student in the same school do not see the same material. The role comes from the session, and it is part of the filter. Access filtering before the model, not after Some systems retrieve broadly and then ask the model to ignore what the user should not see, or filter the answer afterwards. Both are weak. Once a document is in the context, it can influence the answer even if not quoted. A summary can reflect it. A tone can reflect it. A refusal can reveal that it exists. And post-filtering an answer for "leaked content" requires knowing what leaked, which is hard. So the order is fixed: Resolve identity and permissions from the session. Build the retrieval query with mandatory filters for product, tenant and role. Retrieve. Re-check each result's permissions against the session (a cheap second check that catches stale metadata). Only then place results in the prompt. Step 4 sounds redundant. It has caught real inconsistencies in our testing, such as a document whose permissions were changed in the product but whose index metadata had not yet updated. Keeping the index in step with the truth An index is a copy. Copies drift. For RAG to be safe, the index has to follow the source of truth closely. Deletions. When a user deletes a document, a school removes a lesson, or a store closes its account, the corresponding index entries must be removed as part of the same deletion flow, not in a monthly clean-up. We treat the index as one of the places personal and tenant data lives. See Data retention and deletion. Permission changes. If a document moves from "all staff" to "leadership only", the metadata in the index must change too. Until it does, the re-check in step 4 is the safety net. Content updates. For KeşifAtlası, the rules database is the source of truth for visa and relocation eligibility. When a rule changes, the old text must stop being retrieved immediately. Each indexed passage carries the rule identifier and version; retrieval only returns passages whose version matches the current one in the rules database. An out-of-date passage is simply not eligible. See Rules engines for eligibility. We also reindex on a schedule and compare counts between source and index. A mismatch triggers a check. It is a dull control and a useful one. What goes into the index in the first place The safest data in an index is data that was never put there. Before indexing, we ask what the feature actually needs. For a help assistant, it needs help articles, not customer tickets. For a store's policy assistant, it needs policy pages and product information, not order histories with customer names. For EduRelia, it needs lesson content, not student work — and where student work is used, it stays in the student's scope and follows the product's child data rules. See Data minimisation for children in edtech. Where source documents contain personal details the feature does not need — names, emails, phone numbers in a support document, for example — they are removed or masked at ingestion. CastLyra is a good example: talent contact details are governed by explicit access rules on the marketplace, so they are never indexed for AI features at all. The model cannot reveal what it was never given. Prompts, logs and caches Retrieval can be perfect and data can still spread through the surrounding plumbing. Prompts. Retrieved passages are placed into the prompt clearly marked as reference material, with the source identifiers the validator can check. The prompt tells the model to answer only from those passages and to say when they do not contain the answer. This improves quality; it is not our access control. Logs. By default the gateway logs metadata — which documents were retrieved (by identifier), token counts, latency, prompt version — not the passages themselves. Where a product needs full content for evaluation, it is sampled, filtered and kept for a short, defined period in that product's own storage. Caches. A response grounded in tenant or personal documents is either not cached or cached with the tenant and user in the key. Product-level knowledge, such as a help article explanation, can be cached normally. See Latency and caching for AI. Evaluation sets. Test questions and expected answers for a product's RAG feature belong to that product and use synthetic or consented data. They are never assembled from one product's real users to test another. Testing for leakage We test isolation directly, not just answer quality. Cross-tenant probes. In staging, create two tenants with distinctive, made-up documents — a fictional return policy with an unusual phrase, for instance. From tenant A's session, ask questions designed to surface tenant B's phrase. Any hit is a failure. Missing-filter tests. Call the retrieval library without a tenant context and confirm it refuses rather than returning everything. Prompt injection attempts. Ask the assistant, in many phrasings, to ignore its instructions, show all documents or act as another customer. Because scoping does not depend on the prompt, these should fail harmlessly. Deletion tests. Delete a document and confirm it stops appearing in retrieval within the defined time. Version tests. For KeşifAtlası, update a rule and confirm the old passage is no longer returned. These tests run whenever the retrieval library or a product's retrieval configuration changes. A worked example: a school's own materials on EduRelia EduRelia licenses its learning product to schools, which then offer it to students and families. That B2B2C shape — described in B2B2C licensing for schools — makes it a useful test of every rule above. Imagine a school that uploads its own revision notes for a history unit, so that the study assistant can answer questions in the school's own terms. Here is how the retrieval is set up. The school's notes go into that school's own namespace, separate from EduRelia's general lesson content and separate from every other school. Before indexing, the notes are checked for personal details — a teacher's email address at the bottom of a page, a class list pasted by mistake — and anything like that is removed. Each passage is tagged with the school, the unit and the audience: staff only, or staff and students. When a student in that school asks a question, the server reads the school and the role from the session. Retrieval searches two places: EduRelia's general content for the subject, and the school's namespace, filtered to passages marked for students. It never searches another school's namespace, because the library has no way to address it from this session. Staff-only notes, such as a marking guide, are excluded by the role filter and then excluded again by the re-check. If the school ends its licence, its namespace is deleted as part of the offboarding flow, along with any cached answers keyed to it. The general lesson content is untouched. Nothing the school contributed lingers in answers for anyone else. When not to use RAG at all RAG is not always the right tool, and sometimes skipping it is the safest design. If the answer depends on exact, current facts — a store's stock level, an eligibility outcome, a booking status — retrieval over documents is the wrong source. Those facts should come directly from the verified system through a query in code, then be handed to the model to explain. That is the core idea in AI explains, verified data decides. An index of stock levels would be stale within minutes. If the material is small and stable, such as a product's refund policy or a short set of reading guidelines for ZodiVela, it may simply be included in the registered prompt. No index, no retrieval, nothing to filter. And if a feature would need to retrieve personal data to work, we first ask whether the feature is worth that. Sometimes the answer is to redesign it so the personal part is handled by templates and code, with the model only explaining general material. Common mistakes A few patterns worth avoiding: One index "for now". Starting with a single index across products or tenants and planning to split it later. Later rarely comes, and splitting a live index is harder than building two. Admin tools without filters. Internal tools that query the index directly for debugging, bypassing the library. They should use the same enforced path, with least-privilege access. Trusting the model to refuse. "The model is told not to reveal other customers' data" is a hope, not a control. Forgetting derived data. Summaries, embeddings and cached answers are all derived from source documents and must be deleted when the source is. Mixing public and private in one index. Even with filters, a mistake then exposes the private material. Separate them. Summary RAG without data leakage is mostly about the retrieval step, because a model will use whatever it is given. At Oryvelon there is no shared knowledge base: each company has its own sources, indexes, credentials and logs, even when the retrieval code is shared. Within a product, we separate product, tenant and personal knowledge; scope every query with identifiers from the authenticated session rather than the prompt; filter before retrieval and re-check after; keep indexes in step with deletions, permission changes and rule updates; index only what features need; and test isolation directly with cross-tenant probes, missing-filter checks and deletion tests. ## Safety boundaries for consumer AI products: pre and post hooks URL: https://oryvelon.com/insights/ai-safety-hooks-consumer-products/ Category: AI Operations A consumer AI product talks to people you will never meet, in moods you cannot see, about things you did not plan for. Some will test the edges on purpose. Many more will arrive at an edge by accident — a student asking something far outside the lesson, a reader bringing a real worry to what was meant as a light reflection. Safety boundaries for consumer AI are the decisions and mechanisms that keep a product inside what it is for, even when a conversation drifts. This note describes how we set those boundaries for two of Oryvelon's consumer products, ZodiVela and EduRelia, and how we enforce them with pre-request and post-response hooks in our shared AI gateway. It describes our practice. It is not legal advice, and the specific limits for any product depend on its audience and the rules where it operates. Start with what the product is for Most safety problems in consumer AI are really positioning problems. A product that has not decided what it is for will be asked to be everything, and a general-purpose model will try to oblige. So every consumer product at Oryvelon has a short written boundary policy in its source-of-truth document. It answers four questions: Purpose. What is this product for, in one sentence? Audience. Who uses it, and what do we know or not know about them? Never. What must the product never do, regardless of how it is asked? At the edge. What does it do when a conversation reaches a boundary? For ZodiVela, the purpose is personal discovery readings framed for reflection and entertainment. The audience is adults who choose to use it. It must never present readings as prediction, and never give medical, financial, legal or relationship directives. At the edge, it acknowledges the topic, stays with reflection, and points to appropriate real-world help where the situation is serious. See Entertainment positioning for consumer AI. For EduRelia, the purpose is helping students learn, with families and schools involved. The audience includes children. It must never collect more personal information than the task needs, never produce content unsuitable for the student's age, and never become a private confidant that replaces trusted adults. At the edge, it redirects to the learning task and encourages the student to talk to a teacher or parent. The policy is written in plain language, a page or two at most. Everything technical follows from it. Why the prompt is not enough The first instinct is to put the policy in the system prompt. We do, and it matters: the prompt is the main way to shape tone, framing and default behaviour. But a prompt is guidance, not enforcement. Models occasionally ignore instructions, especially over long conversations. Users can phrase requests to slip around them. A prompt change made to improve quality can weaken a boundary without anyone noticing. And a new model version may interpret the same prompt differently. So we treat safety as three layers: Layer What it does Where it lives Positioning Sets expectations for users before any AI is involved Product copy, onboarding, terms Prompt Shapes the model's behaviour and framing Registered, versioned prompt Hooks Enforces limits on every request and response Gateway configuration and product code If the prompt fails, the hooks catch it. If the hooks miss something subtle, positioning means the user was never led to expect advice in the first place. See Onboarding and trust in AI products. Pre-request hooks Pre-request hooks run after the product has built its input and before anything is sent to a model provider. Ours do four jobs. Minimise what is sent The model should receive only what the task needs. For EduRelia, that means no student names, no school names, no contact details and no identifiers beyond a pseudonymous session reference the model never sees in readable form. If a student types their full name and address into a question, the hook removes them before the request leaves. The model gets the maths question, not the child. See Data minimisation for children in edtech. For ZodiVela, a reading needs birth details the user chose to provide and the theme they selected. It does not need their email, payment status or device information, so none of that is included. Classify the request A lightweight classifier, usually a small model or a set of rules, labels each incoming message by topic and risk. Labels are product-specific. ZodiVela cares whether a message concerns health, money, legal matters, a crisis or self-harm, or a request for a definite prediction. EduRelia cares whether a message is on-task, off-task, age-inappropriate, or a sign that a student may need help from an adult. Route before calling the model Some labels change what happens next. A message that suggests a person may be in crisis does not go to the reading or tutoring prompt at all. It goes to a fixed, reviewed response that takes the message seriously and points to real help, written in advance by people, not generated. Requests for predictions about health or money in ZodiVela are sent to a prompt variant that is explicitly reflective and declines to forecast. Enforce limits Per-user rate limits, maximum input length and conversation length limits are applied here. Long conversations are where models drift most, so both products cap them and start fresh sessions rather than letting a single thread run indefinitely. Post-response hooks Post-response hooks run after the model responds and before the user sees anything. Schema and format checks. Where the product expects structured output — a reading with named sections, a tutoring answer with a step list — the response is validated first. Content boundary checks. A second classifier checks the output against the product's "never" list. For ZodiVela, it looks for advice-like directives ("you should leave your job"), certainty about the future ("this will happen"), fatalistic statements, and health or financial claims. For EduRelia, it checks age-appropriateness, off-topic drift and any attempt to ask the student for personal information. Grounding checks. Where the output should reference facts from the input, the hook checks that it does and that it has not introduced new ones. This is the same principle as AI explains, verified data decides, applied to consumer content. Action on failure. If a check fails, the response is not shown. Depending on the configuration, the gateway retries once with a stricter prompt variant, or returns a pre-written, reviewed in-bounds response. It never shows the failed output with a warning attached. Every block is logged by category, without storing the full content by default, so the team can see how often boundaries are being hit and why. A worked example: a ZodiVela reading at the edge Imagine a ZodiVela user who asks for a reading about the coming month and adds: "Should I sell my flat and put everything into a new business? I need to know if it will work." Here is what happens. The pre-request classifier labels the message as a money topic combined with a request for a definite prediction. The product's routing sends it to the reflective variant of the reading prompt, which is written to explore themes such as change, readiness and values without forecasting outcomes or recommending decisions. The model produces a reading. Suppose one sentence slips: "This is a strong month to take a financial leap." The post-response check flags it as a financial directive. The gateway retries once with the stricter variant. The second response stays with reflection: it talks about what the user seems to value, what questions might be worth asking before a big decision, and suggests that a decision like this is best made with someone qualified to look at the numbers. The user receives that second reading. They also see, as with every ZodiVela reading, the standing note that readings are for reflection and entertainment. They are not told that a sentence was blocked. They are simply given a reading that respects what the product is. If the second response had also failed, a pre-written reflective passage for that theme would have been used, and the user's credits would not have been consumed for a personalised reading. See AI fallbacks and degraded modes. A worked example: EduRelia off-task Now imagine a student using EduRelia's study assistant for a science unit. Midway through, they write that they are upset because of problems with friends and ask the assistant what to do. The pre-request classifier labels the message as off-task and personal. EduRelia's policy is clear that the assistant is not a counsellor or confidant. The message is routed to a reviewed response: it acknowledges that the student sounds upset, says kindly that this is something worth talking about with a parent, teacher or another trusted adult, and offers to continue with the science when they are ready. The response is short, warm and does not ask follow-up questions about the situation. It does not collect details. If the message had contained signals of possible risk to the student, the routing would have used a different reviewed response pointing to immediate help, and the product's escalation process would apply as defined with the school. See Human escalation in AI products. Nothing about the conversation is stored beyond what EduRelia's retention rules allow. See Data retention and deletion. Different products, different limits A common mistake in shared infrastructure is a single safety profile applied to everything. It is either too strict for some products or too loose for others. Our gateway provides the hook mechanism; each product configures its own classifiers, labels, thresholds and responses. Compare: ZodiVela EduRelia Audience Adults who opt in Students, including children, with families and schools Core boundary Reflection, not prediction or advice Learning support, age-appropriate, minimal data Streaming Paragraph by paragraph, each checked Generally off; full answer checked first Conversation length Capped per session Tightly capped per session Off-topic handling Gently returns to reflection Redirects to learning and trusted adults Human route Support contact and signposting School and family involvement, defined escalation Our other products have hooks too, tuned to their risks. KeşifAtlası blocks any output that states an eligibility decision not produced by its rules engine. CastLyra checks that generated profile text does not invent credentials and that no contact details appear where marketplace rules do not allow them. MerchNivo checks that suggested actions refer to figures that exist in the store data. Testing the boundaries Boundaries that are not tested decay. We maintain a separate evaluation set for each consumer product's safety behaviour, alongside its quality evaluation set. Each set contains realistic edge cases written by the team: requests for predictions, attempts to get advice, off-topic personal messages, attempts to make the assistant reveal its instructions or pretend to be something else, and inputs containing personal details that should be stripped. For EduRelia, cases are written as a student might actually phrase them, with spelling mistakes and slang. Every prompt change, model change or hook change runs against this set before release. A drop in boundary adherence blocks the release, even if quality scores improve. See Prompt versioning and evaluation. We also review a small sample of blocked and borderline outputs regularly, with personal details removed, to spot new patterns. Test sets use synthetic cases, never real user conversations copied from production. Prompt injection and instruction leaks Consumer products attract people who want to see what the assistant will do if pushed. Some ask it to ignore its instructions. Some paste in text that pretends to be a system message. Some ask it to reveal its prompt, or to role-play as an unrestricted assistant. We assume these attempts will happen and design so they fail harmlessly rather than trying to make them impossible. The most important defence is that nothing sensitive depends on the model obeying. The model does not hold other users' data, because the pre-request hooks never send it. It cannot grant credits, change a subscription or open paid content, because those actions are handled by product code, not by what the model says. Its output is checked by post-response hooks that do not read the user's instructions at all. A user who "convinces" the model to act differently gains, at most, a response that is then blocked and replaced. We also avoid putting anything in the prompt that would be damaging if revealed. Prompts contain framing, tone and structure, not secrets, keys or internal notes. If a prompt leaked, it would be mildly embarrassing, not harmful. Monitoring boundaries in production Safety is not finished at launch. Each consumer product's owner sees a small set of boundary indicators in the weekly review: Block rate by category. How often post-response checks block output, split by reason: advice-like language, prediction, age-appropriateness, personal data, format. Routing rate by label. How often pre-request routing sends messages to reviewed responses, and for which topics. Fallback to reviewed content. How often users receive pre-written content because both attempts failed. User reports. How often users flag a response, and what they flag. A sudden rise in blocks after a prompt or model change usually means the change weakened a boundary, even if it passed evaluation. A slow rise in a routing label, such as more off-task personal messages on EduRelia, may mean the product is being used in ways we did not expect and the positioning or onboarding needs work. Either way, the number starts a conversation rather than waiting for a complaint. When boundaries change Boundaries are not fixed forever. A product may add a new theme, reach a new age group, launch in a new country or respond to new guidance. When that happens, the change starts in the written policy, not in the prompt. The owner updates the boundary policy in the source-of-truth document, noting what changed and why. The classifiers, routing rules and reviewed responses are updated to match. New cases are added to the safety evaluation set before anything goes live. Only then is the prompt adjusted. Doing it in that order keeps the policy, the enforcement and the tests in agreement, and means anyone reading the source of truth later can see why the product behaves as it does. What we avoid A few things we have decided not to do: Over-refusing. A product that refuses anything slightly unusual is safe and useless. Boundaries should be precise, so that normal use feels natural. Lecturing. When a request is outside bounds, a short, kind redirection works better than a paragraph about policy. Hidden personas. Neither product pretends to be a person. Both are clear that the user is talking to an AI. Engagement at any cost. We do not design the assistant to keep people talking. Session caps exist partly for this reason. Collecting more to be safer. More data about a user does not make a product safer. For children especially, less is the safer default. Summary Safety boundaries for consumer AI products work when they start from a clear written policy — purpose, audience, what the product must never do, and what happens at the edge — and are enforced in layers. Positioning sets expectations, the prompt shapes behaviour, and pre-request and post-response hooks enforce limits on every request: minimising input, classifying and routing sensitive topics, checking output against the product's boundaries, and replacing anything out of bounds with reviewed content. ZodiVela and EduRelia share the mechanism but not the rules, because an entertainment product for adults and a learning product for children need different limits, tested continuously. ## Data retention and deletion: schedules, flows, backups and logs URL: https://oryvelon.com/insights/data-retention-and-deletion/ Category: Data, Privacy & Security Most digital businesses keep data because deleting it takes effort and keeping it seems free. It is not free. Every record kept is a record that can leak, be misused, slow down a migration or complicate a customer's request to be forgotten. And most of it is never looked at again. At Oryvelon, each company has its own data retention and deletion practice: a written schedule of what it keeps and for how long, deletion flows that reach every copy, and clear handling of backups and logs. This note explains how we approach it, with examples from across the group. It describes practice, not legal advice. Laws such as GDPR in Europe and the UK and KVKK in Türkiye set expectations about keeping personal data only as long as necessary and honouring deletion requests; children's data rules such as COPPA in the United States add more. Each company takes its own advice on how these apply to it. Why retention is a company decision, not a group one It would be tidy to have one retention policy for the whole group. It would also be wrong. Our companies are different businesses with different users. EduRelia works with students, families and schools, and keeps as little as possible about children. CastLyra is a marketplace where verification records and consent matter. Noveniq is a store with orders, invoices and returns. MerchNivo processes data from merchants' Shopify stores on their behalf. KeşifAtlası handles questions about people's relocation plans. Each has different obligations and different reasons to keep things. So each company owns its retention schedule, stored in its source-of-truth document. What the group shares is the method: the template, the checklist and the review cadence. This follows the broader principle of shared infrastructure, separate data. And because no user data is shared between products, a deletion in one company never needs to be coordinated with another. Building a retention schedule A retention schedule is a table. It does not need to be long, but it needs to be complete. Each row is a type of data, and the columns answer the same questions. Column Question it answers Data type What is it? (Account profile, order record, support message, AI prompt log…) Purpose Why do we keep it? Location Where does it live? Every system, including copies. Retention How long, and from which event? End action Delete, anonymise or aggregate? Owner Who is responsible for it? Mechanism What actually does the deletion? A scheduled job, a flow, manual? Two columns cause most of the work. Location is where hidden copies appear. When a team lists every place an order record lives, they usually find it in the database, the admin tool's export folder, the email service that sent the receipt, the analytics tool, the support tool and a spreadsheet someone made last quarter. All of them belong in the row. Retention needs a starting event, not just a duration. "Two years" means nothing without "from what?" From account closure? From the last order? From the end of the school's licence? The event determines when the clock starts, and it is often the part people get wrong. A simplified, illustrative example for a store like Noveniq might include rows like these. The durations here are placeholders to show the shape, not recommendations: Data type Purpose Retention (illustrative) End action Order and invoice records Fulfilment, returns, accounting As required for accounting, from order date Delete personal fields, keep financial record Customer account Login, order history Until closure, plus a short grace period Delete Support messages Resolving issues A defined period after the ticket closes Delete Marketing consent record Proof of consent While subscribed, plus a period after Delete Raw server logs Security, debugging Short, rolling Delete automatically Short by default Our starting assumption for any new data type is short retention. The team has to justify keeping something longer, and "it might be useful someday" is not a justification. Longer retention needs a concrete reason: an accounting or legal requirement, an active contract, a real product feature that depends on history, or a security need. Where history is useful for analysis, we ask whether aggregated or anonymised data would do. Usually it does. A store does not need five-year-old customer names to know which months are busiest. This is also a cost decision. Storage is cheap per gigabyte, but data that must be protected, backed up, migrated, searched and eventually deleted is not. Deletion flows that reach every copy A deletion request — from a user, a school, a merchant uninstalling an app, or a scheduled job — is only complete when it has reached every place the data lives. We design deletion as a flow with explicit steps, not a single database command. A typical account deletion flow looks like this: Verify the request and confirm the scope (whole account, a specific item, a tenant). Mark for deletion immediately, so the data is no longer shown or used, even before physical removal. Delete from the primary database, including related records. Delete derived data: search indexes, AI retrieval indexes and embeddings, caches, generated summaries. Delete files: uploads, exports, generated documents. Notify connected tools: email platforms, support tools, analytics, payment providers, through their deletion interfaces where they exist. Record the deletion in a log that says what was deleted and when, without keeping the deleted content. Confirm to the requester where appropriate. Step 4 is the one teams forget most often, and it matters more as products use AI. An embedding of a user's document is derived from that document; if the document is deleted and the embedding remains, retrieval can still surface it. See RAG without data leakage. Step 6 is where deletion leaves your own systems. We keep a list, per company, of every external tool that receives personal data and how deletion works in each. If a tool has no way to delete data on request, that is a reason to reconsider using it. Our quarterly review of tools is where that question is asked. Worked examples across the group MerchNivo: when a merchant leaves MerchNivo works with data from a merchant's Shopify store: orders, products, inventory and customer information needed to produce briefings and operations signals. The merchant is the customer; their store's shoppers are people whose data MerchNivo processes on the merchant's behalf. When a merchant uninstalls the app, the store's data is scheduled for deletion after a short, stated grace period, in case the uninstall was a mistake. After that, store data, generated briefings, cached AI outputs and any retrieval index for that store are deleted. Operational metrics that contain no store data — request counts, latency — may remain in aggregate. Shopify's own requirements for apps handling customer data requests are part of the flow as well. See Tenant isolation explained. MerchNivo also avoids copying shopper data it does not need. A refund spike can be described with counts and product names; it does not need the names of the customers who asked for refunds. Less copied data means less to delete. EduRelia: when a school's licence ends For EduRelia, the retention schedule is shorter across the board, and some data is never stored at all. Tutoring conversations are kept only for as long as the product feature and the school agreement require, then deleted. Student profiles hold the minimum needed to run the product. When a school's licence ends, the school's tenant data — student accounts, progress records, uploaded materials, the school's retrieval namespace — is deleted on a schedule agreed with the school, with an export offered first where the school needs it. See B2B2C licensing for schools and Data minimisation for children in edtech. CastLyra: verification records CastLyra has a harder balance. Verification is central to the marketplace's trust, so the fact that a profile was verified, and when, needs to be kept while the profile is active. But the documents used to verify it — identity documents, for instance — are sensitive, and keeping them indefinitely increases risk without adding much. The approach we favour is to keep the verification outcome and date, and delete the underlying documents after a short, defined period once review is complete. When a talent closes their account, their public profile, portfolio and contact details are removed promptly, and consent records are handled according to the schedule. KeşifAtlası: answers about someone's plans KeşifAtlası users answer questions about their nationality, work, family situation and where they hope to move, in order to see which visa or relocation rules may apply. Those answers are personal, and many users only take the free test once. So answers from a free test that does not lead to an account are kept only briefly, long enough to show the result and let the user continue to a paid report if they choose. After that they are deleted, and only aggregated counts remain — how many tests were completed for a given route, for example — with nothing that identifies a person. For paid reports, the answers and the report are kept for a defined period after delivery so the customer can return to it and ask follow-up questions, then deleted unless the customer keeps an account. See From free test to paid report. Backups: expire, don't edit Backups are where deletion gets awkward. A backup is a snapshot; if a user's data was in yesterday's snapshot, deleting it from the live database today does not remove it from the backup. Editing backups to remove individual records is usually impractical and risky. So our approach is: Backups expire on a fixed, documented cycle. Once the cycle passes, data deleted from the live system is gone from backups too. The cycle is part of the retention schedule and part of what we tell users about how long deletion takes to fully complete. Backups are access-restricted, encrypted and used only for recovery, never for analytics or testing. Restores re-apply deletions. If we ever restore from a backup, the deletion log is replayed against the restored data before it goes live, so deleted records do not return. That last point is the one people miss. Without it, a restore quietly undoes every deletion made since the backup was taken. Logs: useful, short-lived and filtered Logs are the other quiet store of personal data. Application logs, server logs, error reports and AI request logs can all contain names, emails, IP addresses or entire user messages. Our rules for logs: Filter at write time. Personal fields are masked before logs are written, not after. Log metadata, not content. For AI features, the gateway records model, prompt version, token counts, latency and outcome by default, not the prompt text or response. See Building an AI gateway. Short rolling retention. Operational logs are kept for a short period, long enough to investigate issues and security events. Separate security logs. Logs needed for security investigations, such as authentication events, may be kept longer, with tighter access. Deletion log without content. The record that a deletion happened says which account and which categories, not what the data was. AI data is data AI features create new places for personal data to live, and each needs a retention rule: Prompt and response logs, if kept for evaluation, are sampled, filtered and stored for a short, defined period in the product's own storage. Caches of model responses expire, and any entry keyed to a user or tenant is deleted with them. See Latency and caching for AI. Retrieval indexes and embeddings follow the source documents they were built from. Evaluation sets use synthetic or explicitly consented data, never copied production conversations. Provider-side retention is checked in each product's AI project settings, so that data sent to a model provider is not kept longer than the product's own policy allows. Deletion versus anonymisation Not every end action has to be deletion. Sometimes a record is useful in aggregate long after the personal part stops being needed. In that case we anonymise: remove or irreversibly transform everything that could identify a person, and keep the rest. It is easy to do this badly. Replacing a name with an internal ID is not anonymisation if the ID can be joined back to the account table. Removing names but keeping exact addresses, birth dates and rare combinations of attributes often leaves people identifiable. Real anonymisation usually means aggregating — counts per week, per product, per region — rather than keeping individual rows with fields blanked out. Our rule of thumb: if the anonymised data would still let someone on the team answer "was this particular person in here?", it is not anonymised enough, and it stays under the retention schedule like any other personal data. Handling individual requests Scheduled deletion covers most data. Individual requests — "please delete my account", "what do you hold about me?" — need a clear path too. Each company publishes how to make a request and handles it through the same deletion flow described above, not through a one-off manual clean-up. The requester's identity is confirmed in proportion to the risk: signing in to the account is usually enough; a request by email for a sensitive account needs more. The team replies within the time the applicable rules set, and says honestly what happens next, including that backups will clear on their normal cycle and that some records, such as invoices, may be kept where the law requires them. For EduRelia, requests about a student usually come through the school or a parent, and the flow reflects that. For MerchNivo, a shopper's request about their data normally reaches the merchant first, and MerchNivo supports the merchant in carrying it out. Reviewing the schedule A retention schedule written once and never checked is just a document. Each company reviews its schedule as part of the operating cadence, typically quarterly, and whenever a new data type, tool or feature is added. The review asks three questions: does every data type we now hold appear in the schedule; are the deletion jobs actually running, with evidence; and has any retention period grown without a reason? We also run a test deletion in staging for a representative account and confirm it disappears from every listed location. Common mistakes No starting event. "Keep for a year" with no defined start is unenforceable. Forgetting derived data. Indexes, caches, exports and summaries outlive their sources. Soft delete forever. Marking records as deleted without ever removing them is hiding, not deleting. Test data from production. Copying real customer data into test environments creates copies no deletion flow knows about. Marketing and product data mixed. Unsubscribing from emails and deleting an account are different requests with different flows. See Marketing consent vs product consent. Summary Data retention and deletion work when each company treats them as design decisions rather than clean-up. At Oryvelon every company owns a written schedule listing each data type, its purpose, every location, a retention period with a starting event, and the mechanism that deletes it. Deletion flows reach every copy — databases, AI indexes and embeddings, caches, files, external tools and logs. Backups expire on a documented cycle and deletions are re-applied after any restore. Logs are filtered and short-lived, AI data is treated as data, and the schedule is reviewed regularly with a real test deletion. ## Tenant isolation explained: separate databases, schemas and row-level security URL: https://oryvelon.com/insights/tenant-isolation-explained/ Category: Data, Privacy & Security Tenant isolation is the answer to one plain question: can one customer's data ever show up for another customer? Not in theory, not "only if someone makes a mistake" — can it happen? In a multi-tenant product, where many customers share the same application, the honest answer depends on how the data is separated, where the separation is enforced and how much of the system sits outside the main database. This article explains the three common tenant isolation patterns — separate databases, separate schemas and row-level isolation — and what each one costs. It also explains how we apply isolation at Oryvelon, where the question has two layers. Our companies must be isolated from each other, and inside each company, every customer must be isolated from every other customer. What a "tenant" actually is A tenant is the unit of ownership in a shared system. In a B2B product it is usually an organisation: a store, a school, an agency. In a consumer product it may be an individual account or a household. The tenant is whoever the data belongs to, from the customer's point of view. Getting this definition right matters more than the technology. For MerchNivo, the tenant is a Shopify store. A merchant who runs two stores has two tenants, because the stores have different products, customers and numbers, and a daily briefing that blended them would be wrong. For EduRelia, the tenant in a licensing deal is the school, and students sit inside it with even tighter rules. For CastLyra, tenants are brand accounts and talent profiles, each with their own visibility and contact-detail rules. Once the tenant is defined, isolation means that every piece of tenant data carries its owner, and the system refuses to hand it to anyone else. Three patterns for tenant isolation Most multi-tenant systems use one of three patterns, or a mix of them. Separate databases per tenant Each tenant gets its own database. The application connects to the right database based on who is signed in. This is the strongest boundary. A query cannot accidentally read another tenant's rows because those rows are not there. Backups, restores, exports and deletion are all per tenant, which makes "delete everything for this customer" a clean operation. Some customers — schools, larger enterprises — will ask for this outright. The cost is operational. Migrations have to run across every database. Connection pooling gets harder. Reporting across tenants, even for your own internal metrics, needs a separate pipeline. With thousands of small tenants, the overhead can outweigh the benefit. Separate schemas per tenant All tenants live in one database server, but each has its own schema: its own set of tables. The application sets the schema for each request. This is the middle ground. The boundary is still structural, and a missing filter will not usually leak data, because the other tenant's tables are in a different namespace. Operations are simpler than one-database-per-tenant, but migrations still multiply and the database's catalogue grows with the number of tenants. Shared tables with row-level isolation All tenants share the same tables. Every row carries a tenant identifier, and every read and write is filtered by it. This is the cheapest and most common pattern for software with many small tenants. It scales well, keeps migrations simple and makes internal reporting easy. It also puts the entire boundary on one thing: the tenant filter being applied every single time. One forgotten WHERE tenant_id = ? in a new report and the boundary is gone. The fix is to stop relying on developers remembering. Databases that support row-level security can enforce the filter themselves, based on a tenant value set for the session. Alternatively, all data access goes through one layer that injects the tenant filter and refuses queries without it. Comparing the patterns Separate database Separate schema Row-level (shared tables) Strength of boundary Strongest Strong Depends on enforcement Cost per tenant Highest Medium Lowest Migrations Per database Per schema Once Per-tenant deletion and export Simple Fairly simple Needs careful queries Cross-tenant internal reporting Hard Moderate Easy (and easy to misuse) Typical fit Few, large or sensitive tenants Tens to hundreds of tenants Many small tenants There is no universally correct choice. The right pattern depends on how many tenants you expect, how sensitive the data is and what your customers are entitled to ask for. Isolation at two levels: why Oryvelon isolates per company and per customer Oryvelon is a company builder. We build, own and operate independent companies, and we share infrastructure between them where it saves time and money. What we do not share is data. We explain the principle in Shared infrastructure, separate data; tenant isolation is how it becomes real. That gives us two levels of isolation. Level one: between companies. Each company has its own data stores, its own credentials and its own access list. MerchNivo's database does not contain EduRelia's tables in a different schema; it is a different store altogether. There is no shared user table across the group and no cross-product identity. This is effectively the "separate database" pattern, applied at the company level, and it is not negotiable. It is also what makes it possible for a company to stand alone — to be sold, spun out or shut down without untangling its data from everyone else's. Level two: between each company's customers. Inside a company, the pattern depends on the product. A product with many small tenants, like a Shopify operations tool, may use shared tables with enforced row-level isolation. A product that serves schools may justify stronger separation for each licensed institution. The source-of-truth document for each company records which pattern it uses and why (see Why every company starts with a source-of-truth document). The two levels answer different threats. Company isolation protects against the group itself becoming a data risk — the temptation to combine audiences, the blast radius of a single breach, the complexity of a merged system. Customer isolation protects against the everyday bug that shows one merchant another merchant's revenue. Where tenant data leaks in practice Most tenant leaks do not come from the main database. They come from everything around it. When we review a product for isolation, these are the places we look first. Caches. A cached response keyed by a URL or a question, but not by tenant, will happily serve store A's answer to store B. Every cache key must include the tenant. Search indexes. A shared full-text or vector index that is filtered after retrieval can still leak through ranking, counts or snippets. Filter at query time, or keep separate indexes. Background jobs. A job queued with a record ID but no tenant context can run with the wrong permissions or write results into the wrong place. Logs and error reports. Stack traces and request logs often contain payloads. If the whole engineering team can read all logs, the logs are a cross-tenant view. Exports and file storage. A predictable file path or a public bucket defeats every database control. AI context. If an AI feature builds its prompt from retrieved documents, the retrieval step is a data-access path and needs the same tenant filter as everything else. We cover this in RAG without data leakage. Analytics. Product analytics that capture free text or identifiers can recreate tenant data in a tool nobody thinks of as a database. A useful rule: derived data belongs to the tenant it came from. Summaries, embeddings, cached answers, aggregates and model outputs all inherit the tenant of their source. AI makes isolation harder, not different AI features add new paths for data to travel, but the principle stays the same. Our shared AI gateway handles routing, budgets, safety hooks and observability for every company. It deliberately does not merge prompts, knowledge bases, keys or users across products. In practice that means: Every AI request carries the company and the tenant it belongs to. Retrieval only searches that tenant's knowledge. Response caches are keyed by company, tenant and prompt version, never just by the question. Observability logs store what is needed to debug and measure — tokens, latency, errors, schema validation results — with tenant content minimised and access restricted. No tenant's data is used to improve another tenant's answers. Take a concrete MerchNivo case. Imagine a store sees a spike in refunds on a single product. The daily briefing flags it and suggests checking the product page and recent supplier batches. The numbers come from that store's data. The explanation may be written by a model. But the model is only ever shown that store's figures, and the cached version of the briefing can only ever be served back to that store. Another merchant with a similar product never sees the signal, and the signal is never "learned" into a shared pool. Enforcing the tenant boundary in code Whatever pattern a product uses, a few engineering habits make the boundary hold. Resolve the tenant once, early. The tenant should be determined at authentication and attached to the request context. Handlers should not accept a tenant ID from a form field or a URL parameter and trust it. Enforce in one place. Either the database enforces row-level rules, or a single data-access layer applies them. Hand-written filters scattered across hundreds of queries will eventually miss one. Fail closed. A query without tenant context should error, not return everything. This is a small design choice that prevents the worst outcomes. Make cross-tenant access explicit. Some internal jobs genuinely need to see across tenants — billing, support tooling, aggregate health checks. Give them a separate, named path with its own permissions and logging, instead of letting them bypass the normal layer. Test for leaks on purpose. Create two synthetic tenants in test, sign in as one, and try to read, update, export, search and prompt your way into the other. Automated tests for this are cheap and catch regressions that code review misses. This depends on having good environments and synthetic test data. Keep secrets per company. Database credentials, storage keys and model provider keys are separate for every company and every environment, so that a leaked key only opens one door. See Secrets management basics. Choosing a pattern: a worked comparison Consider three of our companies side by side. These are illustrations of how we reason, not claims about scale. MerchNivo. Many stores, each with modest data volume, all using the same features. Shared tables with database-enforced row-level isolation fit well. The data is commercially sensitive — revenue, margins, customer orders — so enforcement must be structural and cross-tenant tests run on every release. Store-level export and deletion are designed in from the start, because a merchant who uninstalls should be able to leave cleanly. EduRelia. Schools licensing a product for students is a different risk profile. The data concerns children, the school is typically the data controller, and contracts can require specific handling. Stronger separation per school is easier to explain, easier to audit and simpler to delete at the end of a contract. The extra operational cost is worth paying. We discuss the wider approach in Data minimisation for children in edtech. KeşifAtlası. Users answer questions about nationality, residence and circumstances to check visa and relocation eligibility. Most users are individuals rather than organisations, so each account is its own tenant. The strongest protection here is minimisation: collect only what the verified rules database needs to evaluate eligibility, and keep report data separate from marketing data. The pattern follows the data, the customer and the contract — not a group-wide default. Common mistakes We see the same errors repeatedly in multi-tenant systems, including early prototypes of our own. Treating the admin panel as outside the boundary. Internal tools are often built fast and given broad access. They should respect tenants too, with cross-tenant views limited to the people who need them. Tenant ID in the URL, trusted without checking. Changing /store/123/ to /store/124/ should never work. Shared caches added late. Caching is usually introduced for performance under pressure, which is exactly when the tenant key gets forgotten. Assuming the vector database is "just an index". It holds tenant content in another form and needs the same controls. One set of production credentials for everything. A single key that can read every company's storage turns a small mistake into a group-wide incident. Deletion that only touches the main table. Backups, search indexes, caches and analytics need a deletion story too. See Data retention and deletion. A tenant isolation checklist Before a product goes live, and again when significant features ship, we walk through a short list. The tenant is defined in writing, including edge cases (one person, two stores; one school, many campuses). The chosen isolation pattern is recorded with the reasoning. Tenant context is resolved at authentication and enforced in one layer or by the database. Queries without tenant context fail. Caches, search indexes, file paths and AI retrieval all include the tenant. Logs minimise tenant content and are access-controlled. Cross-tenant internal paths are named, restricted and logged. Automated tests attempt cross-tenant reads, writes, exports and AI queries. Per-tenant export and deletion work end to end. Credentials are per company and per environment. None of this is exotic. It is the discipline of doing the ordinary things every time, in every product. What isolation costs, and why we pay it Isolation is not free. Separate data stores per company mean more infrastructure to run and more migrations to manage. Strict tenant enforcement slows down the quick internal report someone wants on a Friday afternoon. Some growth ideas — "email everyone who uses any of our products" — are simply unavailable to us. We are comfortable with that; the reasoning is in Separating marketing consent from product consent. What we get in return is a portfolio where a problem in one product stays in one product, where customers can be told plainly who holds their data and why, and where each company remains a clean, independent business. For a group whose whole model is building companies that can stand on their own, that trade is easy. Questions to ask any multi-tenant vendor The same thinking works in reverse when we evaluate a tool that will hold our data. A vendor's security page rarely answers the useful questions, so we ask them directly. How is a tenant defined in your system, and can one of our companies be a separate tenant from another? Is the tenant boundary enforced by the database, by a single access layer, or by individual queries? Do caches, search indexes and any AI features respect the same boundary? Can your support staff see our data, and if so, is that access logged and time-limited? Can we export and delete everything for one tenant, including derived data? Is our data used to train or improve features for other customers? Vague answers are an answer in themselves. A vendor that has thought about isolation can usually explain its approach in a few sentences. Summary Tenant isolation is what guarantees that one customer's data never reaches another. The three main patterns — separate databases, separate schemas and row-level isolation on shared tables — trade boundary strength against operating cost, and the right choice depends on the number of tenants, the sensitivity of the data and what customers are entitled to ask for. At Oryvelon we isolate twice: every company has its own data stores with no shared user identity, and inside each company every customer is a tenant with a structurally enforced boundary. Whatever the pattern, the boundary only holds if the tenant is resolved once, enforced in one place, extended to caches, indexes, logs and AI context, and tested deliberately for leaks. ## Child data minimisation in edtech: what to collect, what to refuse and who decides URL: https://oryvelon.com/insights/child-data-minimisation-edtech/ Category: Data, Privacy & Security Child data minimisation in edtech comes down to one discipline: collect what the learning needs, and stop there. It sounds obvious. In practice, almost every product decision pulls the other way. Analytics wants more events. Personalisation wants more profile fields. Sales wants richer usage reports for schools. Growth wants a way to reach families directly. Each request is reasonable on its own, and together they turn a learning tool into a data collection system that happens to teach. EduRelia is Oryvelon's AI learning company for students, families and schools. It carries the strictest data standard in our group. This article explains the practice behind that standard: how we think about schools as controllers, how consent works when the user is a child, what we refuse to collect and how minimisation shapes the AI features themselves. It describes our approach, not legal advice. Laws such as GDPR, KVKK and COPPA set different requirements in different places, and every deployment has to be checked against the rules that actually apply. Why children's data needs a different standard Children cannot meaningfully weigh what they are giving away. They type what they are asked, answer what the screen suggests and rarely read a notice. The data they generate is also unusually long-lived: a learning record, a behaviour note or a written disclosure made at ten years old can follow someone for decades if it is kept and copied. Regulators across many jurisdictions treat children's data as needing extra protection for exactly these reasons. But the stronger argument for an operator is trust. Schools will not adopt a product they cannot explain to parents. Parents will not keep a product they suspect of profiling their child. A company that treats child data carelessly does not have a compliance problem; it has a business problem. So we start from the opposite end. Instead of asking "what data could make this better?", we ask "what is the smallest set of data that lets a child learn well here?" Start from the learning purpose Minimisation only works if the purpose is written down precisely. "Improve learning" is not a purpose; it justifies anything. A usable purpose looks more like this: Show a student the next appropriate exercise in a topic, based on their recent answers. Let a teacher see which students in their class are stuck on which skill. Let a parent see the time spent and topics covered this week. Each purpose points to specific data. The first needs recent answers and a skill model — not a birth date, not a photo, not a home address. The second needs class membership and progress per skill — not the full text of every answer forever. The third needs weekly aggregates — not a keystroke log. When someone proposes a new data field, the question is always "which written purpose needs it?" If none does, the field is not added. If the purpose is new, the purpose is discussed first, and only then the field. Schools as controllers: who decides When a school licenses EduRelia for its pupils — the B2B2C model we describe in B2B2C licensing for schools — the school is usually the party that decides why and how pupil data is used. In GDPR language, the school is typically the controller and the product provider a processor acting on its instructions. Other frameworks use different terms, but the idea is similar: the institution responsible for the child's education decides, and the provider follows. That has practical consequences for how we build. The school's instructions are the scope. If the licence covers maths practice for two year groups, the product processes data for that. It does not quietly add features that use pupil data for something the school did not agree to. The school sees and controls its data. A school administrator should be able to see which pupils are enrolled, what is stored about them, and remove them. When the licence ends, the school's data is returned or deleted according to the agreement. We do not reach around the school. A pupil account created through a school is not a marketing lead. We do not email the child, and we do not use the school relationship to acquire the family as a separate consumer customer without a clear, separate and appropriate process. Our reasoning on this is in Separating marketing consent from product consent. Each school is its own boundary. Schools are isolated from one another, and pupil data from one school is never visible to another. We explain the patterns in Tenant isolation explained. Families can also use EduRelia directly, without a school. In that case the parent or guardian is the account holder, and the relationship is with them. The minimisation rules do not loosen; the only difference is who decides and who receives explanations. Consent that makes sense to families Consent for children is not one checkbox. It is layered, and each layer has a different audience. The institution. In a school deployment, the school agreement sets out what is processed and why. Depending on the jurisdiction and the lawful basis the school relies on, the school may also need to inform or obtain consent from parents. The parent or guardian. For direct consumer use, the parent is the account holder. Where the law requires verifiable parental consent for children under a certain age, the product has to obtain it before collecting personal information from the child. The child. Even when a child cannot legally consent, they deserve an explanation they can understand. A short, plain-language note — "your teacher can see which questions you got right, so they can help you" — respects the child and reduces surprises. We keep a few rules for all layers: Say what happens, not what is permitted. A notice that lists every theoretical use is less honest than one that describes what the product actually does. No bundled consent. Agreeing to use the learning product never includes agreeing to marketing, research or anything else. Withdrawal is as easy as agreement. A parent who wants their child's account deleted should be able to request it without friction. Record it. Every consent or instruction is stored with a date, the text shown and who gave it, within the company that collected it. What we do not collect Case-by-case judgement is unreliable under deadline pressure, so EduRelia keeps a written "do not collect" list in its source-of-truth document. A new feature that needs something on the list needs an explicit, documented decision to change the list, not a quiet exception. Do not collect Why Precise location No learning purpose needs it; it creates physical safety risk. Device contacts, photos or microphone by default Not needed for learning; high sensitivity. Biometric data (face, voice prints, fingerprints) Disproportionate for the purpose and hard to reverse if exposed. Home address and phone numbers for pupils The school or parent holds these; the product does not need them. Full date of birth, where age band or year group suffices Year group supports learning; exact birth dates add identifiability. Free-text "about me" profiles Children disclose personal details in open fields. Advertising identifiers and third-party ad trackers No advertising in a child-facing product. Social features exposing pupils to strangers Contact risk outweighs any learning benefit. Behavioural or emotional profiling Inferences about a child's personality or state are out of scope. This list is deliberately blunt. It is easier to explain to a head teacher, easier to check in a review and harder to erode than a nuanced policy. Minimising inside the data you do collect Minimisation is not only about which fields exist. It also covers precision, retention and access. Precision. Store the year group, not the birth date. Store the skill-level estimate, not every intermediate calculation. Store that a session lasted twenty minutes, not a second-by-second activity trace. Pseudonymisation. Pupils can be identified internally by an ID, with the link to their name held only where the teacher view needs it. Analytics and model evaluation work on IDs, not names. Retention. Answer histories are useful for a period, then summarised or deleted. Class data follows the school year and the licence. Nothing is kept "in case it is useful later". See Data retention and deletion. Access. Teachers see their own classes. School administrators see their school. Support staff see only what they need to resolve a ticket, with access logged. Engineers do not browse production pupil data. The principles are in Least-privilege access for small teams. Testing. Development and staging use synthetic pupils, never copies of real ones. The rule is simple and has no exceptions for "just this once". We explain the setup in Environments and test data. How minimisation shapes AI learning features AI makes minimisation more important, because a model will use whatever context it is given, and a child will type whatever comes to mind. EduRelia's AI features are designed around a narrow context. When a student asks for help with a problem, the model receives the problem, the relevant curriculum material and a small summary of the student's recent performance on that skill. It does not receive the student's name, school or anything else it does not need to explain the step. Several safeguards run through our AI gateway: Input filtering. If a child types personal details — a phone number, an address, a full name — into a free-text box, the system can detect and strip them before the text reaches a model provider or a log. Safety hooks. Messages that suggest a child may be at risk are not treated as ordinary learning questions. They follow a defined route agreed with the school, rather than being answered by a model. We describe the general approach in AI safety hooks for consumer products and Human escalation in AI products. No training on pupil data. Student inputs are not used to train or fine-tune models, and model providers are configured not to retain or train on the data where that option is available. Structured outputs. Explanations are generated against a schema, and curriculum facts come from verified material, not from the model's memory. AI explains; verified material decides. Minimal logs. Observability records latency, cost, errors and validation results. Conversation content is minimised, retained briefly and access-restricted. One concrete example. Imagine a student writing, mid-exercise, "I can't do this, my dad says I'm stupid." A general chatbot might offer comfort and carry on. That message is not a maths question. In a school deployment it should follow the school's safeguarding route, and the product should not store it in a general conversation log where it could be seen out of context or kept indefinitely. Designing for that moment is part of minimisation: knowing what not to keep, and who should see what is kept. What schools and parents should be able to see Minimisation is more convincing when it is visible. We aim for EduRelia to be able to give any school or parent: A plain list of the data fields stored about a pupil, and why each exists. The retention period for each category. A description of which staff roles can access what. A statement of which model providers process data, in general terms, and under what restrictions. A working process to export or delete a pupil's data. If we cannot produce one of these for a feature, the feature is not ready for children. Common mistakes in edtech data practice Patterns we deliberately avoid: Collecting now, deciding later. "We might use it for research" is the most common reason for over-collection. Research needs its own purpose, basis and approval. Consumer analytics in a child product. Generic analytics tools capture far more than a learning product needs, often including device identifiers and cross-site tracking. We keep analytics minimal and first-party where possible. Treating the school relationship as a funnel. Pupils enrolled by a school are not a consumer audience. Open text everywhere. Every free-text field is an invitation to disclose. Use structured answers where they serve learning just as well. Copying production into test. One developer's laptop becomes the least protected copy of children's data in the company. Forgetting derived data. Summaries, embeddings and cached explanations built from pupil work are pupil data too. A minimisation review for new features Before any EduRelia feature that touches pupil data ships, it passes a short written review. Which written purpose does this feature serve? What new data does it collect, and at what precision? Is anything on the "do not collect" list involved? Who can access the data, and how is access logged? How long is it kept, and how is it deleted? Does any of it reach a model provider, and with what restrictions? How will a school or parent see it and remove it? What happens if a child types something sensitive into it? The review takes minutes for most features. For the few where it takes longer, that time is well spent. One pupil's data, from enrolment to deletion It helps to follow a single hypothetical pupil through the whole lifecycle, because minimisation decisions show up at every stage. Enrolment. The school's administrator adds a class. For each pupil, the product needs a display name the teacher recognises, a year group and a class link. It does not ask for a birth date, a home address, a parent's phone number or a photo. If the school's own systems hold those, they stay there. First sessions. The pupil signs in with a method the school chooses. The product records answers to exercises and builds a skill estimate per topic. It does not record device location, and it does not load third-party trackers. Ongoing use. After a period, detailed answer histories are condensed into skill summaries. The teacher still sees who is stuck on fractions; nobody needs every wrong answer from six months ago. Moving class or school. Moving between classes in the same school is a change the school makes. A pupil who leaves the school leaves its data behind; nothing follows them to a new institution unless the schools and family arrange that themselves. End of licence. When the agreement ends, the school's data is exported if requested and then deleted, including derived summaries and cached explanations, within the period set in the contract. At no point does the pupil become a marketing contact, a research subject or a data point in another product. That is what minimisation looks like when it runs end to end. Why this matters for the group EduRelia's standard is the strictest in Oryvelon, but it influences everyone. It proves that shared infrastructure can carry very different data rules for different companies, which is the whole point of sharing infrastructure while keeping data separate. It also sets a reference point. When another company asks whether it needs a data field, "would we collect this for a child?" is a useful test even when the user is an adult. Summary Child data minimisation in edtech means collecting only what a specific, written learning purpose needs, at the lowest useful precision, kept for the shortest useful time and seen by the fewest people. When schools license a product, the school usually decides how pupil data is used and the provider follows its instructions; families using the product directly are the decision-makers for their own children. At EduRelia we pair that with layered, unbundled consent, a blunt written list of data we do not collect, AI features that receive minimal context and never train on pupil data, and a short review for every feature that touches children's information — practices that make the product easier to trust, easier to explain and easier to adopt. ## Marketing consent vs product consent: why we keep them separate in every company URL: https://oryvelon.com/insights/marketing-consent-vs-product-consent/ Category: Data, Privacy & Security Marketing consent and product consent are two different things, and most privacy trouble in growing companies begins when they are treated as one. A person who signs up to use a product has agreed to let that product work for them. They have not agreed to receive promotions, to be added to a newsletter, or to hear from a sister brand they have never heard of. At Oryvelon we build and operate several independent companies. That structure creates an obvious temptation: pool the audiences, cross-promote everything, and turn every sign-up into a group-wide lead. We decided early not to do that. This article explains the distinction between product consent and marketing consent, how we record consent per company, why we do not build cross-brand marketing lists, and the practical details that make the rules hold. It describes our practice, not legal advice; GDPR, KVKK, electronic marketing rules and other laws set specific requirements that each company checks for its own markets. Two different questions The simplest way to see the difference is to look at the question each one answers. Product consent — or, more precisely, the basis on which a product processes someone's data — answers: what does this service need in order to do what you asked? A Shopify merchant installing MerchNivo needs the product to read store orders so it can write a daily briefing. A visitor taking KeşifAtlası's eligibility test needs the product to process their answers to evaluate them against the rules database. Often this is not "consent" in the legal sense at all; it is processing necessary to provide the service. Either way, it is scoped to the service. Marketing consent answers: may we contact you to promote things? It is optional. It is about the relationship after or beyond the service. And critically, refusing it must not stop the person from using the product. When the two are bundled — "By creating an account you agree to receive offers" — the person has no real choice. Many regulators consider that invalid consent for marketing. Even where the law is looser, it is a poor way to start a relationship. What goes wrong when they are merged Merging product and marketing consent causes a predictable chain of problems. The list is inflated and unreliable. Every user becomes a "subscriber", but many never wanted to be. Open rates fall, complaints rise and deliverability suffers across the whole sending domain. Withdrawal becomes confusing. If unsubscribing from emails seems to mean closing the account, people complain instead, or mark messages as spam. Product messages get mixed with promotions. A password reset, a failed payment notice and a discount campaign should not share one channel and one preference. When they do, people unsubscribe from everything and then miss critical service messages. Nobody can prove anything. When a person asks "why am I getting this?", a merged system has no record of what they actually agreed to. For a group of companies, add one more: a merged list crosses brand boundaries. Someone who bought a phone case from Noveniq did not ask to hear about visa eligibility reports or beauty courses. How we separate them in each company Every Oryvelon company follows the same pattern, adapted to its product. Service messages are part of the product. Account confirmations, receipts, security alerts, report delivery, school notices and similar messages are sent because the service needs them. They are not promotional, and they do not carry sales offers. People cannot switch off security alerts, but they can control optional product notifications. Marketing is a separate, unticked choice. Sign-up forms show marketing as its own option, not pre-selected, with plain wording about what the person will receive and from whom. The product works the same whether they tick it or not. Marketing is branded clearly. A newsletter from ZodiVela comes from ZodiVela, about ZodiVela. The light "An Oryvelon company" endorsement may appear in a footer, but it is not an invitation to send group-wide promotions. We explain the brand structure in Naming, domains and brand architecture. Withdrawal is one click. Every marketing message has a working unsubscribe. The preference centre, where one exists, separates marketing topics from service notifications. Consent can be narrower than "all marketing". Where it helps the person, we split consent by channel (email, SMS) or topic. More granular choices lead to fewer blanket unsubscribes. Per-company consent records The most important structural decision is where consent lives. At Oryvelon, each company holds its own consent records, in its own data store, under its own access controls. There is no group-level consent table, just as there is no group-level user table. This follows from our wider principle of shared infrastructure, separate data and from the isolation patterns in Tenant isolation explained. A consent record is more than a boolean. A useful record includes: Field Example (hypothetical) Subject Internal ID for the person in this company Purpose Marketing email: product updates and offers Channel Email Status Granted / withdrawn Source Checkout page, newsletter form, account settings Wording shown The exact text or a version reference to it Timestamp When it was given Withdrawal timestamp When it was withdrawn, if it was Company The company that collected it — and the only one that may use it Two details are easy to miss. First, store the wording version, not just "yes". If the text changes, you need to know what each person actually agreed to. Second, keep the history. A person who subscribed, unsubscribed and subscribed again has three events, not one flag. When a company uses an email or messaging platform, the platform receives only the contacts and preferences for that company, in a workspace or account that belongs to that company. Credentials for those platforms are per company too (see Secrets management basics). No cross-brand marketing lists This is the rule people outside the group most often question, so it is worth stating plainly: we do not build marketing lists that span our brands, and consent given to one company is never used by another. It would be possible to ask for group-wide consent — "hear from other Oryvelon companies" — and some groups do. We choose not to, for four reasons. It does not match what people expect. Our companies serve very different people for very different reasons. A parent using EduRelia, a talent on CastLyra and a merchant on MerchNivo do not expect to be treated as one audience. For EduRelia, student data is never used for marketing at all. It weakens each company's independence. Every company should be able to stand alone. A company whose audience depends on a shared group list cannot be sold, spun out or evaluated on its own merits. A company whose list it built itself is a real asset. It blurs the evidence. When we decide whether to continue, stop or scale a company, we want to know whether it can acquire its own customers. Borrowed audiences hide the answer. It creates a group-level data risk. A single cross-brand list would be the most valuable and most exposed dataset in the group. Not having it is the simplest protection. What we do instead is ordinary, honest marketing per company. If a CastLyra user wants to know about other Oryvelon companies, the companies page is public. If a company wants a partnership with a sister company — say, a Sinem Keser Beauty Academy course promoted on its own channels and linking to content elsewhere — it is done through public content, not by exchanging contacts. Wording that is clear on the page Consent quality starts with the words on the form. A few patterns we use and avoid: Use: - "Send me product updates and offers from ZodiVela by email. You can unsubscribe at any time." - "Email me when new eligibility rules affect my report" — a service-adjacent option that is still optional and clearly scoped. Avoid: - Pre-ticked boxes. - "Partners", "affiliates" or "group companies" as vague recipients. - Consent hidden in terms of service. - Double negatives: "Untick this box if you do not wish to not receive…" - Making the marketing checkbox look like a required field. For product surfaces with AI, the same honesty applies to onboarding: say what the product does with the person's data before they give it. We cover that in AI product onboarding and trust. Worked examples across our companies The same rule plays out differently in each business. Noveniq (DTC commerce). A customer checking out needs to receive order confirmations and delivery updates; that is part of the purchase. Whether they want new-product emails is a separate checkbox at checkout, unticked. If they decline, they still get their order updates. Noveniq's list contains only people who chose Noveniq marketing. MerchNivo (software for Shopify stores). The merchant installs the app and receives daily briefings; that is the product. MerchNivo may also offer an optional newsletter about new features or e-commerce operations. The merchant's store data — their customers, their orders — is never a marketing asset for MerchNivo. The shoppers in a merchant's store are the merchant's customers, not ours, and they never enter any MerchNivo list. KeşifAtlası (eligibility reports). A visitor takes a free test and may buy a paid report; we describe the flow in From free test to paid report. Emailing the report and notifying the buyer when a relevant rule changes can be part of what they purchased. Promotional email about other destinations is a separate choice. The answers they gave about their nationality and circumstances are never used for targeting. CastLyra (marketplace). Talent and brands receive messages about bookings, verification and account security as part of using the marketplace. Promotional emails about features or events are optional. Talent contact details are governed by the marketplace's access rules and are never turned into a marketing list for anyone, inside or outside the group. EduRelia (edtech). Pupils enrolled by a school are never marketed to. Families who use EduRelia directly may opt in to updates as adults; the child's account is not a marketing channel. See Data minimisation for children in edtech. Consent and measurement Marketing consent interacts with analytics and advertising in ways that are easy to get wrong. Cookie and tracking consent is its own question. Where a company needs consent for analytics or advertising cookies, that is recorded separately from email marketing consent. Someone can accept one and refuse the other. Advertising audiences follow the same rules. Uploading customer lists to an advertising platform is a use of personal data. It needs its own basis, it stays within the company that holds the data, and it never mixes lists across brands. Campaign tagging is not personal data. Consistent UTM parameters tell us which campaign brought a visit without needing to identify anyone. See Analytics without surveillance. Consent as events, not flags A common implementation stores marketing consent as a single yes/no column on the user record. It is simple, and it loses almost everything you need later. We prefer to treat consent as an append-only series of events. Each event says what happened: granted, withdrawn, re-granted, changed channel. The current state is derived from the latest event for each purpose and channel. This design has a few practical benefits: It answers "why am I getting this?" Support can show exactly when and where someone opted in, and what the form said. It survives wording changes. When a company rewrites its opt-in text, old consents keep their original version reference. If the new wording covers something materially different, the company knows whose consent needs refreshing. It makes syncing safer. When a withdrawal happens in the email platform (through its unsubscribe link), it is written back as an event in the company's own store. The company's record stays the source of truth, not the vendor's. It supports deletion properly. When a person asks to be erased, the company deletes their personal data but may keep a minimal suppression entry so they are not re-added by a later import, where the applicable rules allow that. See Data retention and deletion. The model is small — one table, a handful of columns — and it pays for itself the first time someone asks a difficult question. When a company is sold, closed or spun out Because each Oryvelon company is designed to be independent, we plan for the day one changes hands or stops. Per-company consent makes this straightforward. If a company is sold, its consent records go with it, together with the wording people agreed to, and the buyer can see exactly what it is inheriting. No other company's contacts are mixed in, so nothing has to be untangled. If a company is closed, its lists are handled according to its own retention rules, rather than quietly absorbed by a sister brand. People who signed up for a closed brand are not "migrated" to another one. Our reasoning on endings appears in Continue, stop or scale. Merged, group-wide consent makes each of these moments harder and riskier. Separate consent makes them routine. Common mistakes A short list of errors we check for when reviewing a company's forms and flows: A marketing checkbox that is pre-ticked, or required to proceed. A "welcome series" of promotional emails sent to everyone who created an account. Unsubscribe links that lead to a login screen. A single "email preferences" toggle that turns off security alerts along with offers. Consent stored in the email platform only, with no record in the company's own system. Importing an old list without knowing what those people agreed to. A new brand launched by "warming up" an existing brand's audience. A quick consent review checklist When a company launches or changes a form, we run through these questions: Is the service usable without marketing consent? Is marketing consent unticked, separate and clearly worded? Does the wording name the specific brand? Is the consent recorded with source, wording version and timestamp in the company's own store? Do service messages avoid promotional content? Does every marketing message have a one-click unsubscribe that works without logging in? Is withdrawal reflected everywhere the list is used within a reasonable time? Is there any path by which this list reaches another company in the group? (The answer must be no.) The cost, stated honestly Keeping consent separate and per company means smaller lists and slower early growth. A new company cannot start with a ready-made audience. Some campaigns that would be easy for a merged group are unavailable to us. We accept that. Each list we have is one that people chose, for a brand they know, and it will hold up to scrutiny from customers, partners and regulators. It also gives us honest signals about which companies are genuinely earning attention. That is worth more to a company builder than a large list of reluctant contacts. Summary Product consent and marketing consent answer different questions: what a service needs to work for someone, and whether that person wants promotion. Keeping them separate means service messages stay free of offers, marketing is an unticked and clearly worded choice, withdrawal is one click, and every consent is recorded with its wording, source and date. At Oryvelon each company keeps its own consent records, consent never transfers between brands, and there are no cross-brand marketing lists. It makes growth slower at the start, but every audience in the group is one that its company earned and can defend. ## Least-privilege access for small teams: roles, contractors, 2FA and access reviews URL: https://oryvelon.com/insights/least-privilege-access-for-small-teams/ Category: Data, Privacy & Security Least-privilege access sounds like something for large companies with security departments. In practice it matters most in small teams, where one person's laptop, password or bad afternoon can reach everything. A five-person company where everyone is an admin on every system has no real boundaries. It just has not been tested yet. At Oryvelon we operate several independent companies with small teams and a mix of employees, founders and contractors. Least privilege and two-factor authentication are group-wide principles. This article explains the practical model we use — roles, contractor accounts, 2FA, quarterly access reviews and offboarding — in a form any small team can adopt without buying anything new. What least privilege means in a small company Least privilege is simple to state: every person and every system gets the minimum access needed for the work they are doing now, for as long as they are doing it. Three parts of that sentence carry the weight. Minimum. A designer needs the design files and the staging site, not the payment dashboard. A support agent needs to look up a customer's order, not export the whole customer table. Now. Access granted for a project six months ago is not automatically needed today. As long as. Temporary work should mean temporary access, with an end date. Small teams resist this because it adds friction. Someone asks for access, someone else has to grant it, and in the meantime work waits. The trick is to make the right access easy to grant and the wrong access easy to notice, rather than to make everything slow. Why "everyone is admin" breaks down It is worth being specific about what goes wrong, because the risks are ordinary rather than dramatic. Compromised accounts. A phishing email or a reused password turns one person's account into an attacker's account. If that person had access to everything, so does the attacker. Honest mistakes. Admin rights make it possible to delete the wrong project, change the wrong DNS record or run a script against production instead of staging. Narrower permissions turn many of these mistakes into error messages. Departures. When someone leaves, the team has to find every system they could touch. With universal admin, that is every system, often including ones nobody remembers setting up. Data exposure. People who can browse production data will, occasionally, for good reasons: debugging, answering a support question, checking a hunch. Each time, personal data is seen by someone who might not have needed to see it. Tenant and company boundaries. In a group like ours, one account with access to every company's systems undermines the whole design of shared infrastructure with separate data. The boundary between companies is only as strong as the accounts that can cross it. Start with roles, not people The most useful single step is to define roles per company and grant access through them. A role is a named bundle of permissions tied to a kind of work. For a typical Oryvelon company, a starting set looks like this: Role Typical access Production personal data Owner Billing, domains, account recovery, all admin settings Only when needed, logged Engineer Code, preview and staging environments, deploy pipeline No direct browsing; break-glass only Operator / support Admin tools scoped to support tasks Limited to the records in a ticket Marketing Website CMS, analytics, the company's own email platform Aggregates and its own consented list only Contractor (scoped) The specific repositories, files or tools for one project None unless the project requires it Read-only Dashboards and reports None A few rules make roles work: Few roles, clearly named. Six roles that everyone understands beat twenty that nobody can explain. Per company. Roles are defined inside each company. Being an engineer on MerchNivo does not grant anything on CastLyra. Where someone genuinely works across companies, they hold separate roles in each, reviewed separately. Access through roles, not favours. When someone needs something outside their role, either their role changes or they get a time-limited exception that is written down. "I'll just make you admin for now" is how permanent admins are created. Owners are few. Owner-level access — billing, domain registrars, account recovery, the ability to add other admins — sits with a very small number of people, with the strongest authentication. Losing control of a domain registrar account is one of the worst things that can happen to a small company; we treat it accordingly in Domain, DNS and email security. Production data is the most sensitive privilege In most small companies, the most dangerous permission is not "admin" in a dashboard. It is the ability to read production personal data directly: the database console, the raw export, the log viewer with full payloads. We treat that access as rare and deliberate. Engineers work in non-production environments by default, with synthetic data. See Environments and test data. Support tools show what support needs. A support view for KeşifAtlası might show the status of a report and the purchase record, without exposing every answer the person gave about their circumstances unless the ticket requires it. Break-glass access exists but is visible. When an incident genuinely requires direct access, it is granted to a named person, for a limited time, with a note of why, and the access is logged. Stricter companies get stricter rules. For EduRelia, where the data concerns children, direct access to pupil data is even more limited. The reasoning is in Data minimisation for children in edtech. Contractors: named, scoped, dated Contractors are normal in small companies and in a group like ours, where specialist work — a design sprint, a migration, a security review — does not justify a full-time hire. Contractor access is also where least privilege most often fails, because the work is urgent and the relationship is temporary. Our rules for contractor access: Named accounts only. Every contractor has their own account in every system they use. No shared logins, no "use the team account". One company, one job. Access is scoped to the company and project they were engaged for. A contractor building a landing page for Noveniq does not get access to Noveniq's order data, let alone to another company's systems. End date on day one. The access expiry is set when access is granted, matching the engagement. Extending it is a deliberate decision. 2FA required. No exceptions for "it's just for two weeks". No production personal data by default. If the project truly needs it, that need is written down and the access is limited to the minimum. Their own tools stay theirs. Contractors should not copy company data into personal drives, personal AI tools or personal accounts. We say this explicitly in the engagement. A named internal owner. Someone on the team is responsible for each contractor's access and for removing it. The same logic applies to agencies and service providers. Within the group, WeAreMedia is the services company; when it works for a sister company, it receives the same scoped, named access as any external provider, not a shortcut because it shares an owner. Two-factor authentication everywhere Mandatory two-factor authentication is the highest-value security control a small team can adopt. It turns a stolen password from a breach into an inconvenience. Our practice: Required on every system that supports it — email, code hosting, cloud and hosting consoles, domain registrars, payment dashboards, analytics, e-commerce admin, design tools, password managers. Authenticator apps or hardware keys over SMS where the system allows. SMS codes are better than nothing but vulnerable to SIM swapping. Hardware keys for owner accounts where possible — registrars, primary email, billing. Recovery codes stored safely, not in the same place as the password and not in a chat message. Systems without 2FA are flagged. If a tool cannot support it, we ask whether we should be using that tool at all, or at least restrict who has an account. This is part of our quarterly tool review. A shared password manager makes all of this practical. Each company has its own vault or collection, shared only with the people in the relevant role. Service accounts and API keys are people too Least privilege is not only about humans. Integrations, scripts, CI pipelines and AI features all hold credentials, and they often have more access than any person on the team. Apply the same rules: Each integration gets its own credential, scoped to what it needs. A script that reads orders should not be able to issue refunds. Credentials are per company and per environment, so a leaked staging key does not open production and a MerchNivo key does not open CastLyra. Keys have owners and are rotated on a schedule and when people leave. Keys live in a secrets manager or environment configuration, never in code, tickets or chat. We go into detail in Secrets management basics. The same principle drives our AI gateway: it holds separate model provider keys per company and enforces per-company budgets, so no product can spend another product's allowance or see its traffic. The quarterly access review Access drifts. People change projects, contractors finish, tools get added in a hurry. A quarterly access review catches the drift before it turns into risk. It fits naturally into the operating cadence we already run for each company. A practical review for a small company takes an hour or two: List the systems. Start from the company's tool inventory: code hosting, hosting, database, email, domain registrar, payment, analytics, marketing, support, design, AI providers, password manager. Export or view each system's user list. Who has an account, and with what role? Check each account against current reality. Is the person still working on this company? Is their role still right? Is 2FA on? Check service accounts and keys. Does each have an owner? Is it still used? When was it last rotated? Remove or reduce. Anything unexplained is removed. Anything broader than necessary is narrowed. Record the review. A short note: date, reviewer, systems checked, changes made. Future reviews start from it. The first review is usually uncomfortable — old accounts, forgotten integrations, a former contractor still on the design tool. That is the point. By the third review it becomes routine. Offboarding on the same day When someone leaves, or a contractor finishes, access should end the same day. In a small team, offboarding usually fails not because nobody cares but because nobody has a list. Our offboarding checklist, per company: Disable the person's primary identity (email or single sign-on) first, which cuts off many downstream systems. Remove their accounts from every system in the tool inventory, including ones outside single sign-on. Remove them from the password manager and shared vaults. Rotate any shared secrets they could have seen, and any keys they created. Transfer ownership of files, repositories, automations and scheduled jobs they owned. Revoke personal access tokens and OAuth grants they authorised to third-party apps. Check domain registrar, DNS and email settings for any access or forwarding they set up. Record the offboarding with the date and the systems covered. For someone who worked across several Oryvelon companies, the checklist runs once per company. Separate systems mean separate removals — more work, but no single forgotten account can reach everything. Onboarding: get access right on day one Most access problems start at onboarding, when a new person needs to be productive quickly and the easiest path is to copy someone else's permissions. Copying permissions copies every exception that person accumulated. A better routine: Start from the role, not a colleague. Assign the role that matches the job and add nothing else. Create accounts from the company's identity, so they can be disabled centrally later. Turn on 2FA before granting access to anything sensitive. Give a short written note of what the person has access to and how to request more. People ask properly when they know how. Record the grant in the same place the access review will look. Twenty minutes of care on the first day saves hours of archaeology at the first review. AI tools and least privilege AI tools add a new category of access that many teams have not thought through. An assistant connected to email, a drive, a code host or a database acts with the permissions of the account that connected it. If that account is an admin, the assistant effectively is too. Our rules are simple. Connect AI tools with the narrowest account and scopes that work. Do not paste production personal data, credentials or customer records into AI tools that are not approved for that company's data. Product AI features go through the company's own configuration in the shared gateway, with that company's keys and limits, rather than through personal accounts. And when someone leaves, their AI tool connections are revoked along with everything else. A worked example Imagine CastLyra hires a contractor for six weeks to improve the talent verification flow. Here is how least privilege shapes the engagement. On day one, the internal owner creates a named account for the contractor on the code host, scoped to the CastLyra repositories involved, and on the staging environment. 2FA is enabled before the first commit. The access expiry is set to the end of week six. The contractor works against synthetic talent profiles in staging; they never see real talent contact details, identity documents or verification results. In week four, the contractor needs to understand a verification failure that only happens in production. Instead of granting database access, an engineer on the team reproduces the case with synthetic data, or extracts the relevant non-personal facts — error codes, timings — for the contractor. At the end of week six, the contractor's accounts expire automatically. The internal owner confirms removal, checks that no keys were created under the contractor's name, and notes the offboarding. At the next quarterly review, the contractor does not appear anywhere. Nothing about this is heavy. It is a few deliberate steps that keep verified talent data where it belongs; see Verification in talent marketplaces for why that data is so sensitive. Common mistakes Shared "team" logins for tools that charge per seat. The savings rarely justify losing accountability. Founders as permanent admins everywhere without 2FA hardware keys. The most powerful accounts deserve the strongest protection. Personal email addresses as account owners. When that person leaves, the company may lose the ability to recover the account. Integrations authorised by individuals and never re-homed. When the person leaves and their account is disabled, the integration breaks — or worse, keeps working under an account nobody controls. Access reviews that only check the main systems. The forgotten analytics tool or old staging server is usually where problems hide. Granting production access to speed up debugging and never taking it back. Summary Least-privilege access for small teams means giving each person and system the minimum access needed for their current work, for only as long as they need it. In practice that comes down to a few roles per company, production personal data treated as a rare and logged privilege, contractors with named, scoped and time-limited accounts, mandatory two-factor authentication everywhere, per-company and per-environment credentials, a quarterly access review and a same-day offboarding checklist. At Oryvelon we run these practices separately in every company, so that access boundaries match data boundaries and no single account can reach the whole group. ## Secrets management basics: API keys, rotation and keeping credentials out of code URL: https://oryvelon.com/insights/secrets-management-basics/ Category: Data, Privacy & Security Secrets management is not glamorous, and most small companies postpone it until something goes wrong. That is a mistake, because a leaked API key is among the most common and most avoidable security incidents there is. A key pasted into a chat, committed to a repository, or bundled into a website's JavaScript can be copied in seconds and misused for months. At Oryvelon we run several independent companies on shared infrastructure. Shared infrastructure only works if credentials stay strictly separate, so secrets management is a group-wide standard rather than something each team improvises. This article covers the basics we apply everywhere: what counts as a secret, why keys stay server-side, why every company and environment gets its own, how rotation works, and the rules that keep secrets out of repositories and conversations. What counts as a secret A secret is any value that grants access to something. If someone who had the value could read data, spend money, send messages or impersonate a system, it is a secret. Typical secrets in a small digital company include: API keys for model providers, payment processors, email platforms and other services Database connection strings and passwords Access tokens for platforms such as Shopify apps, advertising accounts and analytics APIs Webhook signing secrets Session and token signing keys Encryption keys Cloud storage credentials SSH keys and deploy keys Recovery codes for important accounts Some values look harmless but are not. A "test" key for a payment provider may still expose account details. A read-only analytics token may still reveal revenue. When in doubt, treat it as a secret. It helps to think of secrets as physical keys. You would not photocopy the office key and leave copies on desks, send a photo of it in a group chat, or tape it to the front door. Most secrets leaks are the digital version of one of those three things. Rule one: secrets stay on the server Anything shipped to a browser or a mobile app can be read by anyone who uses it. Minified JavaScript is not hiding anything; browser developer tools show every network request and every bundled value. So the first rule is simple: keys that grant access to paid services or private data never leave the server. If the frontend needs something that requires a key, it calls our own server, and the server makes the call with the key. This matters especially for AI features. A model provider key embedded in a web page can be extracted and used to run someone else's workloads at our expense. At Oryvelon, every AI call goes through our AI gateway on the server side. The gateway holds the provider keys, enforces per-company budgets, applies safety hooks and logs usage. The browser never sees a provider key, and a product cannot spend another product's budget. We explain why that budget control matters in Cost discipline for AI products. Some keys are designed to be public — a publishable payment key, a site key for a bot check, a public analytics measurement ID. These are fine in frontend code because the provider expects them to be visible and restricts what they can do. Read the provider's documentation; if it does not say a key is safe to publish, assume it is not. Rule two: separate secrets per company Oryvelon's companies share tooling, patterns and some infrastructure. They do not share credentials. MerchNivo has its own database credentials, its own model provider keys, its own email platform account and its own storage keys. So do CastLyra, KeşifAtlası, EduRelia and every other company. There is no group-wide key that opens several companies' data. The reasons are practical: Blast radius. If one company's key leaks, one company is affected. The incident does not become a group-wide one. Independence. A company that can stand alone must be able to take its credentials with it if it is sold or spun out, without breaking anyone else. Clear costs. Per-company keys make it easy to see what each company spends on each provider, which feeds straight into unit economics per product. Data boundaries. Separate credentials are one of the mechanical guarantees behind shared infrastructure, separate data and tenant isolation. The same applies to third parties we connect. A merchant who installs MerchNivo grants a Shopify access token for their store. That token is stored for that store only, used only for that store's features, and removed when the merchant uninstalls. Rule three: separate secrets per environment Every company runs several environments: local development, preview builds, staging and production. We describe the setup in Environments and test data. Each environment gets its own secrets. Environment Typical secrets Can it reach production data? Local Test-mode provider keys, local database No Preview Test-mode keys, preview database with synthetic data No Staging Staging keys, staging database with synthetic data No Production Live keys, production database Yes, by design The point is that a non-production key should be useless against production. If a developer's laptop is lost, or a preview build leaks its configuration, the attacker gets access to synthetic data and test-mode services. Nothing real. This also prevents a quieter class of mistakes: a script run locally that was meant for test data but was pointing at production because someone copied the live connection string "just to check something". Where secrets should live Secrets need a home that is access-controlled, auditable and separate from code. For a small team, that usually means two places. A secrets store or the hosting platform's encrypted environment configuration for application secrets. The application reads secrets at runtime from the environment. Access to view or change them is limited to the people who deploy that company's systems. A team password manager for human credentials: dashboard logins, registrar accounts, recovery codes. Each company has its own vault or collection, shared only with the relevant roles. Where secrets should never live: In the code repository, including private repositories and "temporary" config files In chat messages, direct messages or group channels In tickets, task descriptions or project documents In emails In screenshots and screen recordings In AI assistant conversations In spreadsheets In logs and error reports Private repositories and private channels are not safe places. They are copied, forked, backed up, searched and connected to other tools. Former team members and connected apps may still have access. A secret in a chat history is a secret in every device and integration that has synced that history. Keeping secrets out of repositories Accidental commits are the classic leak. A few habits prevent most of them. Use an example config file (for instance a .env.example) that lists the required variable names with placeholder values, and keep the real file out of version control with an ignore rule. Turn on secret scanning in the code host and, ideally, a pre-commit hook that blocks commits containing strings that look like keys. Never "fix it in the next commit". Once a secret is committed, it is in the history. Removing it later does not remove it from clones, forks or caches. The secret must be rotated. Review diffs before pushing, especially configuration changes. Keep build logs clean. CI output should mask secret values, and build steps should not print the environment. Scoping: give each key the least it needs Least privilege applies to keys as much as to people; see Least-privilege access for small teams. Use restricted keys where providers offer them. A key that only needs to read orders should not be able to create refunds. A key for sending transactional email should not manage the account. One key per integration. If the website, the support tool and a nightly job all need the same provider, give each its own key. When one leaks or is retired, the others keep working. Set limits. Where providers support spending caps, rate limits, IP restrictions or allowed domains, use them. A model provider key with a monthly cap turns a leak from a large bill into a small one. Give every key an owner and a description. "Created by whom, for what, in which company and environment" should be answerable for every secret. Rotation Rotation means replacing a secret with a new one and retiring the old one. Keys should be rotated: On a schedule, so that any silent leak has a limited lifetime. When someone with access leaves, especially if they could have seen the value. Immediately after any suspected exposure. When a vendor or integration is retired. Rotation is painless only if it is designed in. That means applications read secrets from configuration rather than hard-coding them, deployments can pick up new values without code changes, and — where the provider allows — two keys can be valid at once during the switchover so there is no downtime. We track rotation in the same inventory as ownership. The quarterly tool review and the quarterly access review are natural moments to check for keys that are overdue, unused or unowned. When a key leaks: a short runbook Leaks happen. What matters is how quickly the key stops working. Our runbook is short on purpose. Revoke or rotate the key now. Do not wait to investigate first. A revoked key cannot be misused further. Deploy the replacement to the systems that legitimately need it. Check the provider's logs for unexpected use since the likely exposure time: unusual volume, unknown IP addresses, unexpected resources. Remove the value from where it appeared — the message, the commit history, the ticket — understanding that this is cleanup, not protection. Assess impact. Did the key expose personal data? If so, follow the company's incident process, which may include notification obligations under the applicable law. Fix the cause. Why was the key there? Add the scanner, change the process, update the checklist. Write it down. A brief note of what happened and what changed. The single most important instruction is the first: revoke before you investigate. Keep a secrets inventory You cannot rotate, scope or revoke what you do not know exists. Every Oryvelon company keeps a simple inventory of its secrets. It does not contain the values themselves — those stay in the secrets store — only the facts about each one. Field Why it matters Name and provider So anyone can recognise it Company and environment Confirms it is not shared across boundaries Purpose What breaks if it is revoked Owner Who is responsible for it Scope and limits What it can do and how much it can spend Where it is used Which services read it Created and last rotated When rotation is due A spreadsheet is enough for a small company, provided it holds no secret values. The inventory turns a leak from a scramble ("where else do we use this?") into a checklist, and it makes the quarterly review fast. Secrets in local development Developers need credentials to run software locally, and this is where good intentions often slip. The safest pattern is that local development only ever uses local or test-mode secrets: a local database, sandbox keys for payment and email providers, a low-limit development key for model providers. Developers should never need production secrets on their machines. If a bug only reproduces in production, the fix is better logging, a reproduction with synthetic data in staging, or a controlled break-glass session — not a copy of the live environment file on a laptop. Laptops are lost, synced to personal cloud storage and shared with other software far more often than servers are. Secrets and AI assistants AI coding assistants and chat tools have become a new route for secrets to leak. A developer pastes an error message that includes a connection string. Someone asks an assistant to "fix this config" and includes the whole environment file. An assistant with access to a repository reads a committed key and repeats it in an answer. Our rules: do not paste secrets into AI tools, redact configuration before sharing it, and treat any secret that has been pasted into an external tool as exposed and due for rotation. Connected assistants get the narrowest access that works, as with any other integration. Webhooks and inbound secrets Secrets management is not only about keys we use to call others. It also covers secrets that let us trust incoming requests. Webhooks — notifications that a platform sends to our server when something happens, such as an order being created or a payment succeeding — usually come with a signing secret. Our server should verify that signature before acting on the request. Without it, anyone who discovers the webhook URL can send fake events. For MerchNivo, that means verifying that a webhook genuinely comes from Shopify for a specific store before it updates that store's signals. For Noveniq, it means payment notifications are verified before an order is marked as paid. Verified inputs are part of the same principle we apply to AI: decisions come from data we can trust, as described in AI explains, verified data decides. A worked example: adding a model provider to a new company Imagine KeşifAtlası is adding an AI explanation feature to its eligibility reports. Here is how secrets management shapes the work. The company creates its own account with the model provider — not a shared group account — and generates separate keys for staging and production. Each key is restricted to the capabilities the feature needs and given a spending limit that matches the company's budget. The keys are stored in the company's environment configuration and registered with the AI gateway under KeşifAtlası's configuration, alongside the prompt versions and budgets for that company. The frontend calls KeşifAtlası's own server, which calls the gateway, which calls the provider. The browser never sees a provider key. Staging uses the staging key against synthetic applicant profiles. Production uses the production key. The inventory records both keys: owner, purpose, environment, creation date, rotation due date. At the next quarterly review, someone checks they are still in use and not overdue. None of this takes long. It just has to be the default rather than an afterthought. Common mistakes Putting a model provider key in frontend code "for the prototype" and forgetting it is there at launch. Using one production key across several products because it was quicker to set up. Sharing the production database password in a chat so a colleague can "quickly check something". Copying the production environment file to a laptop to reproduce a bug. Keys created under a personal account that nobody else can manage. No spending limits on usage-billed services. Deleting a leaked secret from a commit or message and considering the problem solved. Summary Secrets management basics come down to a handful of firm rules. Treat every credential that grants access as a secret. Keep keys that open paid services or private data on the server, never in browsers, apps, repositories, tickets, chat or AI conversations. Give each company and each environment its own secrets so a leak stays small, scope every key to the least it needs, give it an owner and limits, rotate it on a schedule and after any exposure, and revoke first when something leaks. At Oryvelon these rules are group-wide because they are what make shared infrastructure safe: the tools are shared, the credentials never are. ## Environments and test data: local, preview, staging and production without real personal data URL: https://oryvelon.com/insights/environments-and-test-data/ Category: Data, Privacy & Security Environments and test data are where good privacy intentions most often quietly fail. A company can have careful access controls, clear consent records and strict retention policies in production — and then a developer copies the production database to a laptop to reproduce a bug, and every one of those controls stops applying. At Oryvelon we build and operate several independent companies, each with its own product and data. We run the same environment model in all of them: local, preview, staging and production, with production personal data allowed in production only. This article explains what each environment is for, how we build synthetic test data that is good enough to replace real data, and the habits that keep the boundary intact. Why environments matter for privacy, not just engineering Environments are usually discussed as an engineering concern: catching bugs before release, testing migrations, previewing designs. All true. But from a data point of view, every environment is also a place where data lives, and every place where data lives needs access control, secrets, logging and deletion. If only production has real personal data, you only have to get those things right once, in the environment with the strongest controls. If staging, preview builds and laptops have copies too, you have to get them right four times — and you usually do not. There is a simpler argument too. People trust a company with their data for the purpose of using its product. They did not agree to be a test fixture. For EduRelia, where the data concerns children, this is non-negotiable. For every other company, it is still the right default. The four environments Each Oryvelon company runs four kinds of environment. Smaller or earlier-stage companies may combine preview and staging at first, but the principles are the same. Environment Purpose Data Integrations Who has access Local Development on a developer's machine Synthetic seed data Sandboxes and mocks The developer Preview A temporary build for each change or branch Synthetic seed data Sandboxes Team and scoped contractors Staging Production-like rehearsal before release Larger synthetic dataset Sandboxes, test-mode keys Team and scoped contractors Production The live product Real customer data Live integrations Minimal, role-based, logged Local Local is where engineers write and run code on their own machine. It should be fast to set up, easy to reset and impossible to damage anything real from. A single command should create a local database and fill it with synthetic data. Preview Preview environments are short-lived builds, often one per change, used to review a feature before it is merged. Designers, product people and contractors can click through them. Because they are created often and shared widely, they must never contain real data or live credentials. Staging Staging is the dress rehearsal. It should match production as closely as practical — same configuration shape, same migrations, similar data volume — so that performance problems and migration failures show up before release. It uses a larger synthetic dataset and test-mode integrations. Production Production is the live product. It holds real data, uses live integrations and has the strictest access. Engineers do not work in production day to day; they deploy to it. Direct access to production data is rare, deliberate and logged, as described in Least-privilege access for small teams. Separate secrets for every environment An environment is only separate if its credentials are separate. Each environment in each company has its own database, its own storage and its own keys. Test-mode keys for payment providers, sandbox accounts for email platforms, low-limit development keys for model providers. The rule that matters: a non-production secret must be useless against production. If a preview build's configuration leaks, the attacker should find synthetic data and sandbox services. We cover this in more depth in Secrets management basics. This also prevents the classic accident — a script meant for local data that runs against production because someone had the live connection string in their environment. The core rule: no production personal data in test We state it plainly in every company's source-of-truth document: production personal data does not leave production. Not as a database dump, not as a CSV for a demo, not as "just a few rows" to reproduce a bug. The rule has no casual exceptions, because casual exceptions are how the rule dies. The tempting cases, and what we do instead: "I need to reproduce a customer's bug." Reproduce the shape of the problem with synthetic data. Use logs, error codes and non-personal identifiers to understand what happened. If direct access is genuinely required, it happens in production, through a controlled break-glass process, not by copying the data out. "The demo looks empty." Build a better demo dataset. Realistic synthetic data is better for demos anyway, because it can be curated to show the product well. "We need realistic volume for a performance test." Generate it. Synthetic data can be produced at any scale. "We'll anonymise the copy." Anonymising real data is harder than it looks. Removing names and emails often leaves enough in addresses, timestamps, free text and combinations of fields to re-identify people. Generating data that was never real avoids the problem entirely. What good synthetic test data looks like The rule only survives if the alternative is good. Poor test data — five users called "Test User", three orders, no edge cases — pushes people back towards production copies. So we treat synthetic data as a product in its own right. It is realistic in shape. Names, addresses, amounts, dates and text follow the formats and distributions of real data. A synthetic Shopify store for MerchNivo has products with variants, orders with discounts and partial refunds, returning and new customers, and several months of history. It includes the edge cases. The most valuable test data is the data that breaks things: very long names, names in several scripts, zero-value orders, refunds larger than the original order due to a data error, missing fields, time zones around midnight, a school with one pupil, a talent profile with no photos. Every production bug that turns out to be data-shaped becomes a new synthetic case. It is obviously fake. Synthetic emails use reserved example domains. Phone numbers use ranges that cannot reach anyone. Names can be realistic but should not be drawn from real customer lists. Nobody should ever wonder whether a record is a real person. It is seeded automatically. A script creates the dataset from scratch. Local, preview and staging environments are populated by the same scripts, so everyone works from the same known state. It is versioned with the code. When the schema changes, the seed data changes in the same change. Test data that falls out of step with the code is quickly abandoned. It covers tenants. Every dataset includes at least two tenants, so cross-tenant isolation can be tested on every run. See Tenant isolation explained. Synthetic data for each kind of company Our companies are different businesses, and their test data reflects that. Some examples of what each company's synthetic dataset needs to show. MerchNivo. Several synthetic stores with different sizes and product mixes. Seasonal patterns, a sudden spike in refunds for one product, a stockout, a discount code that is being overused, a store with almost no data. These let the team check that the operations signals fire correctly and that the daily briefing explains them accurately. The numbers in a briefing always come from store data; in test, that "store data" is synthetic but realistic. CastLyra. Synthetic talent profiles at different verification stages, brands with different permissions, bookings in various states and contact-detail access that should and should not be granted. Profile images are generated or clearly placeholder, never real people's photos. KeşifAtlası. Synthetic applicant profiles across nationalities, residence situations and document states, designed to exercise each branch of the verified rules database. When a rule changes, a synthetic profile that should flip from eligible to not eligible is added so the change can be tested before release. We describe the engine in Rules engines for eligibility. EduRelia. Synthetic schools, classes and pupils with varied progress patterns. No real children, ever — including in demos for schools. The wider reasoning is in Data minimisation for children in edtech. Noveniq. A test-mode store with synthetic products, customers and orders, plus sandbox payment and shipping integrations. Because Noveniq is also a real store we use as a testbed, the distinction matters: what we learn from the live store comes from production analytics and operations, handled under production rules; what developers test against is synthetic. Testing AI features without real data AI features raise the stakes, because testing them well needs a lot of realistic input, and prompts are easy to paste real data into. We build evaluation sets from synthetic inputs: synthetic store snapshots for MerchNivo briefings, synthetic applicant scenarios for KeşifAtlası explanations, synthetic student questions for EduRelia. Each case has an expected structure and, where relevant, expected facts. These sets run whenever a prompt version or model route changes, as we describe in Prompt versioning and evaluation. The staging configuration of our AI gateway uses separate keys and budgets from production. Logs from staging contain only synthetic content, so they can be inspected freely while debugging. When production behaviour reveals a problem — an explanation that misreads a pattern, for example — we turn it into a synthetic evaluation case that reproduces the pattern, rather than copying the real input into the test set. Promoting code, not data A useful mental model: code moves up through environments; data does not. Changes start locally, get a preview build, pass through staging and are deployed to production. Each step runs the same migrations and the same code. Data never travels the other way. Production data does not flow down to staging, preview or local. Configuration moves up with the code, but secrets do not — each environment reads its own secrets from its own configuration. This keeps deployments repeatable and environments independent. Access to each environment Different environments call for different access. We scope them accordingly. Local belongs to the developer. Because it holds only synthetic data, a lost laptop is an inconvenience, not a data incident. Preview and staging are accessible to the team working on the company and to scoped contractors. Because they hold no real data, a contractor can be productive on day one without being granted anything sensitive. Production access is role-based, minimal and logged, per company. This is one of the underrated benefits of synthetic data: it makes collaboration easier. A designer, a contractor or a new engineer can work in staging immediately, without an access request for personal data and without a risk review. Rehearsing migrations in staging Database migrations are one of the main reasons staging exists. A migration that adds a column in a second on a small local database may lock a large production table for minutes, or fail halfway on a row that violates a new constraint. Staging catches most of these problems if its data is shaped like production. That is why staging's synthetic dataset is larger than local's and includes the awkward records: nulls where the schema used to allow them, very old rows created before a field existed, tenants with unusually large histories. We run every migration against staging first, time it and check that the application still works with both the old and new shape during deployment. When staging passes a migration that later fails in production, the lesson goes back into the seed data. The failing pattern becomes a synthetic record, so the same class of problem cannot surprise us twice. Resetting environments on purpose Non-production environments get messy. Test accounts pile up, half-finished records linger, someone changes a setting by hand and forgets. Over time, staging stops matching the seed data and bugs appear that nobody can reproduce. The fix is to make resets cheap and routine. Preview environments are created and destroyed with each change. Staging can be rebuilt from migrations and seed scripts on demand, and is reset on a regular rhythm. Local environments reset with one command. Because nothing in them is real, a reset loses nothing that matters, which is exactly the freedom synthetic data is meant to provide. Handling support cases without copying data Support is where the pressure to copy real data is strongest. A customer reports that their dashboard shows the wrong total, and the fastest route seems to be pulling their records into a test environment. We take a different route. Support staff describe the problem using non-personal details: the tenant identifier, the feature, the time, the error code, the shape of the data ("an order with two partial refunds and a currency change"). An engineer recreates that shape with synthetic records in staging and fixes the bug there. If the investigation truly needs the real records, it happens inside production, through the support tooling or a logged break-glass session, and nothing is exported. For a MerchNivo merchant, this means their store's figures stay in production even while their bug is being fixed. Common mistakes "Temporary" production dumps that live on a laptop or shared drive indefinitely. Staging connected to live integrations, so test orders trigger real emails or real payments. Shared credentials across environments, so staging can reach production storage. Stale seed data that no longer matches the schema, which pushes people back to real data. Screenshots of production in tickets and design files, exposing real customer details to anyone with access to the project tool. Test data with real-looking email addresses at real domains, which can send mail to strangers. Analytics and error tracking in test environments pointing at the same projects as production, mixing test and real data in reports. A setup checklist for a new company When we start a new company, part of the technical groundwork is setting up environments properly from the first week. Create separate local, preview, staging and production configurations. Create separate databases, storage and secrets for each. Use sandbox or test-mode accounts for every third-party integration outside production. Write the seed script before the first feature, with at least two tenants. Add edge cases to the seed data as they are discovered. Point analytics and error tracking in non-production to separate projects, or disable them. Record the rule "production personal data does not leave production" in the source-of-truth document. Restrict production access to named roles, with logging. Doing this at the start takes a day or two. Retrofitting it after a year of production copies takes much longer, and some of those copies are never fully found. Summary A clean environment model gives each stage of development a clear job: local for building, preview for reviewing, staging for rehearsing and production for real customers. Each environment has its own secrets and its own data, and only production holds real personal data. That rule only survives if synthetic test data is good — realistic, full of edge cases, obviously fake, automatically seeded and versioned with the code — so we invest in it as seriously as in the product. Across Oryvelon's companies, from MerchNivo's synthetic stores to EduRelia's synthetic classrooms, this keeps customer data in the one place designed to protect it, and makes every other environment safe to share, break and rebuild. ## Domain, DNS and email security for a multi-brand group URL: https://oryvelon.com/insights/domain-dns-and-email-security/ Category: Data, Privacy & Security Most security incidents at small companies are not clever. A domain expires because the card on file changed. A former contractor still has the registrar login. Somebody deletes a DNS record while "tidying up" and the checkout stops sending order emails. A stranger sends invoices that appear to come from the company, and customers pay them. None of this requires a sophisticated attacker. It requires a gap. For a group like Oryvelon, which builds and runs several independent companies, those gaps multiply with every brand. Eight companies, each with a primary domain, a few defensive registrations, a handful of email services and a DNS zone, is a lot of surface to leave to memory. This article describes our baseline for domain, DNS and email security: how we register and hold domains, how we keep DNS consistent, and how we set up SPF, DKIM and DMARC for each sending domain. None of it is exotic. The value is in doing all of it, for every brand. Why domains are infrastructure, not admin A domain feels like a one-off purchase. In practice it is the root of almost everything a digital company depends on. The website resolves through it. Email is delivered to and from it. Password resets for dozens of services go to addresses on it. Payment providers, app stores and ad platforms verify ownership through it. Search engines tie a site's reputation to it. Lose control of a domain, even for a few hours, and you can lose all of those at once. Someone who controls your DNS can point your site elsewhere, receive your email, issue certificates for your name and reset passwords on accounts registered to your addresses. That is why domain handling sits in our infrastructure standards alongside access control and secrets, not in a spreadsheet of subscriptions. We treat the registrar account as one of the most sensitive systems the group has. The naming and brand architecture work decides which domains a company needs. This article is about keeping them. Registration and ownership The first rule is simple: domains belong to the group, in a group-controlled account. Not in a founder's personal account, not in a freelancer's account, not in whichever account happened to be open when the name was bought. This matters more than it sounds. When a domain sits in an individual's account, the company's most important asset is tied to one person's email, one person's phone for two-factor codes and one person's continued goodwill. When that person leaves, changes phone numbers or simply forgets which registrar they used, the company inherits a problem. Our registration baseline looks like this: Group-controlled registrar account, with two-factor authentication on every login that can reach it. Named access. A small number of named people, each with their own login. No shared passwords. See least-privilege access for small teams. Registrant contact details that point to a role mailbox the group controls, not a personal inbox. If the registrar needs to reach us about a transfer or a dispute, the message must arrive. A domain register. One written list of every domain the group holds, which company it belongs to, what it is used for (primary, redirect, defensive, email only) and when it renews. The register is a small thing that solves a surprising number of problems. When a company like KeşifAtlası needs to know whether it owns a regional variant of its name, the answer is one lookup. When we review costs, defensive domains that no longer earn their keep are visible. When a company is ever sold or spun out, the list of what goes with it already exists. That last point connects to a wider principle: every company should be able to stand alone, and its domains are part of what it takes with it. Locks, auto-renew and expiry Two settings prevent most domain disasters: the registrar lock and auto-renew. Registrar lock (often called transfer lock, shown as clientTransferProhibited in WHOIS or RDAP data) stops the domain being transferred to another registrar unless the lock is first removed from inside the account. It is free and turns a common hijacking route into something that requires access to our account first. For the most important domains, some registries offer a stronger registry-level lock that requires manual verification before changes. For a primary brand domain, that extra step is often worth considering. Auto-renew should be on for every domain we intend to keep. Renewal failures are rarely about money. They happen because a card expires, a billing email goes to someone who left, or a renewal notice lands in spam. To make auto-renew actually work, we also: keep a valid payment method on the account and review it when cards change; renew primary domains for multiple years where it makes sense, so a single failed charge is not a crisis; send registrar notifications to a monitored role address; check upcoming renewals as part of our regular operating cadence. The quarterly review is where defensive domains get a decision. If a typo domain has sat unused for two years and nothing points to it, we may let it lapse on purpose. The important word is on purpose. An expiry should be a choice written in the register, never a surprise. One DNS standard for every brand Each company has its own DNS zone. What they share is a standard: the same structure, the same naming conventions, the same baseline records and the same rules about who can change what. A shared DNS standard is not about uniformity for its own sake. It is about making anomalies visible. When every zone follows the same pattern, a record that does not belong stands out immediately. When every zone is different, nobody can tell whether an odd CNAME is a forgotten experiment, a legitimate service or something worse. Our DNS standard covers: Area What the standard says Hosting of zones Zones are managed in a group-controlled DNS account with two-factor authentication and named access. Apex and www One canonical host per brand; the other redirects to it over HTTPS. Mail records MX, SPF, DKIM and DMARC follow the email baseline below. Verification records TXT records for service verification are labelled in the change log and removed when the service is dropped. Subdomains Created only for a named purpose; every subdomain has an owner and a reason. CAA Restricts which certificate authorities may issue certificates for the domain. TTLs Sensible defaults; lowered temporarily before planned migrations, then restored. Changes Every change is recorded: what, who, when and why. Two items on that list deserve a note. Dangling records are the quiet risk. A subdomain points to a hosting service or storage bucket that was later deleted. If someone else can claim that resource name on the same service, they can serve content on your subdomain. The fix is housekeeping: when a service is decommissioned, its DNS records go with it, and periodic reviews look for records pointing at things that no longer exist. CAA records limit which certificate authorities can issue certificates for a domain. They do not stop every attack, but they narrow the ways someone could obtain a valid certificate for one of our names, and they cost nothing to publish. Where the registry and DNS host support it, DNSSEC adds signatures that protect against forged DNS answers. It needs care, because a misconfigured DNSSEC setup can take a domain offline, so we enable it deliberately and check it after any change of DNS host. Email authentication: SPF, DKIM and DMARC Email was designed without any built-in proof of who sent a message. Anyone can put any address in the From line. SPF, DKIM and DMARC are the three standards that, together, let receiving mail servers check whether a message claiming to come from a domain really did. For a group with several brands, this matters twice. It protects customers from convincing phishing in our names. And it protects deliverability: mailbox providers increasingly expect bulk senders to authenticate properly, and messages that fail these checks are more likely to be filtered or rejected. SPF: which servers may send Sender Policy Framework is a TXT record on the domain listing the servers and services allowed to send mail for it. A typical record looks like this: v=spf1 include:_spf.example-mailservice.com include:mail.example-helpdesk.com -all The includes name each legitimate sending service. The final -all tells receivers that anything else is not authorised. SPF has practical limits worth knowing. It checks the envelope sender (the technical return path), not the visible From address that people actually read. It breaks when mail is forwarded. And a record may trigger at most ten DNS lookups; stack too many includes and the record fails for everyone. That limit is one reason we keep a separate list of every service that sends as each domain, so the SPF record reflects reality rather than history. DKIM: signing each message DomainKeys Identified Mail adds a cryptographic signature to each outgoing message. The sending service holds a private key; the domain publishes the matching public key in DNS under a selector, for example s1._domainkey.brand.com. Receivers use it to verify that the message was signed by someone authorised and was not altered in transit. Each sending service should sign with the brand's own domain, not only with the service's default domain. That detail is what makes DMARC alignment possible. We also prefer keys of adequate length and rotate them when a service changes or on a periodic schedule. DMARC: the policy that ties them together Domain-based Message Authentication, Reporting and Conformance is the layer that tells receivers what to do when a message fails, and asks them to send reports. A DMARC record lives at _dmarc.brand.com: v=DMARC1; p=none; rua=mailto:dmarc-reports@brand.com; adkim=s; aspf=r DMARC introduces alignment: the domain in the visible From address must match the domain that passed SPF or DKIM. This closes the gap SPF leaves on its own. A message can pass SPF for some unrelated domain and still fail DMARC because it does not align with the brand it claims to be from. The policy (p=) has three values: none — monitor only; deliver as normal but send reports. quarantine — treat failing mail as suspicious, usually sending it to spam. reject — refuse failing mail outright. Rolling DMARC out without breaking your own email The common mistake is jumping straight to p=reject and discovering, a week later, that the helpdesk's replies, the invoicing tool and the review-request service were all sending unaligned mail. Customers stop getting messages and nobody notices until someone complains. We move each sending domain through stages: Inventory. List every service that sends as the domain: the mailbox provider, the store platform, transactional email, marketing email, helpdesk, invoicing, form notifications. This is the step people skip, and it is the one that matters most. Authenticate. Add each service to SPF and set up DKIM signing with the brand's domain. Monitor. Publish DMARC at p=none with aggregate reporting. Read the reports for a few weeks. They show every source sending as the domain, legitimate or not. Fix. Anything legitimate that fails alignment gets fixed. Anything unexpected gets investigated. Enforce gradually. Move to p=quarantine, optionally with pct= to apply it to a fraction of mail first, then to p=reject once reports are clean. Keep watching. New services get added over time. Reports keep telling us when something new starts sending. The pace depends on the company. A brand that sends little email can reach enforcement quickly. A commerce company like Noveniq, where order confirmations, shipping updates and marketing all come from different systems, takes longer, because every one of those systems needs to align before we enforce. Domains that never send email A frequent gap in multi-brand groups is the domain nobody thinks of as an email domain. Redirect domains, defensive typos, regional variants and retired brand names all have one thing in common: if they publish no email records, anyone can try to send mail in their name, and some receivers will accept it. For every domain that should never send mail, we publish: v=spf1 -all v=DMARC1; p=reject; and, where appropriate, a null MX record signalling that the domain accepts no mail. Three records, a few minutes of work, and a whole class of lookalike phishing becomes harder. In a group where a customer could plausibly receive mail from several related brands, this closes an easy gap. Separate sending domains, separate reputations Our principle of shared infrastructure, separate data has a direct email equivalent: shared standards, separate sending identities. Each company sends as its own domain. MerchNivo sends product and account email as MerchNivo. EduRelia sends as EduRelia. WeAreMedia sends as WeAreMedia. No company sends customer email from an Oryvelon address, and no company sends on behalf of another. This does three useful things: Reputation stays separate. If one brand's marketing email draws complaints, it does not drag down another brand's transactional delivery. Customers are not confused. A parent who signed up for EduRelia should never receive a message about a Shopify tool. Separate sending domains reinforce the house-of-brands structure and our rule of no cross-product user data. Each company can stand alone. If a company is ever separated from the group, its email identity goes with it intact. Within a single company, we often go one step further and separate marketing from transactional mail, for example by sending newsletters from a subdomain such as news.brand.com. That keeps password resets and order confirmations flowing even if a campaign performs badly. It also makes consent management cleaner, which we cover in marketing consent vs product consent. Access, credentials and change control All of the above can be undone by one weak login. The registrar account, the DNS account and the mailbox administration console are among the systems we protect most tightly: Two-factor authentication on every account, preferably with an authenticator app or hardware key rather than SMS. Named, individual access. No shared logins. When someone's role changes, their access changes the same day. Least privilege. Most people who need to update a website never need registrar access at all. API tokens scoped narrowly, stored as secrets and rotated. A token that can only edit one zone is far safer than one that can edit every zone. See secrets management basics. Recovery paths checked. Recovery email addresses and phone numbers on these accounts belong to the group, not to former team members. Changes to DNS follow the same discipline as code changes, on a smaller scale. We write down what is changing and why before we change it, keep a record of the previous value, and check the result afterwards. For planned migrations, TTLs are lowered in advance so a rollback propagates quickly. A checklist for a new company When we start a new company, the domain and email baseline is part of the launch checklist. It usually takes less than a day, and it is far easier to do before launch than to retrofit later. Register the primary domain and any defensive variants in the group registrar account. Turn on registrar lock and auto-renew. Record the domains in the register with their owner company and purpose. Create the DNS zone following the standard: canonical host, redirects, CAA, HTTPS everywhere. List every service that will send email as the domain. Publish SPF covering exactly those services, and set up DKIM on each. Publish DMARC at p=none with reporting to a monitored address. Publish v=spf1 -all and p=reject DMARC on every domain that will not send mail. Schedule the DMARC review: read reports, fix alignment, move towards enforcement. Confirm two-factor authentication and named access on the registrar, DNS and mail admin accounts. Add the domain's renewal date and DMARC status to the portfolio review. The phase-zero site usually goes live on top of this baseline, so the very first page a visitor sees already sits on a locked, authenticated domain. Common mistakes we design against A short list of the failures we have seen elsewhere and built our process to avoid: Domains bought on a personal card in a personal account, then stranded when that person leaves. Auto-renew on, payment method expired. The setting is only as good as the card behind it. Two SPF records on one domain. Only one is allowed; two means both are invalid. SPF records that grow forever, still listing services dropped years ago and edging past the ten-lookup limit. DMARC stuck at p=none. Monitoring is a stage, not a destination. A policy of none protects nobody. Forgotten verification and CNAME records that outlive the services they were created for. Everyone has admin. Convenience today, incident tomorrow. Summary Domain, DNS and email security is unglamorous work that prevents some of the most damaging and most avoidable incidents a small company can face. Our baseline for every Oryvelon company is consistent: domains owned by the group, locked and set to auto-renew, recorded in one register; DNS zones that follow one written standard so anomalies stand out; SPF, DKIM and DMARC on every sending domain, rolled out from monitoring to enforcement based on what the reports show; locked-down records on every domain that should never send mail; and separate sending identities so each brand keeps its own reputation and can stand alone. It takes little time per brand and removes a whole category of risk across the group. ## What an AI e-commerce employee is, and what it is not URL: https://oryvelon.com/insights/ai-ecommerce-employee/ Category: Commerce & SaaS "AI employee" is a phrase that invites both too much and too little. Too much, because it suggests something that runs a business while the owner sleeps. Too little, because plenty of products with that label are a chat window that answers questions about a dashboard nobody opened. MerchNivo is the Oryvelon company building an AI e-commerce employee for Shopify stores. We chose that description deliberately, and we spend a lot of time on what it does and does not mean. This article sets that out: the job the software does, the parts it is not allowed to do, and why we think the boundary between suggestion and approval is the most important design decision in the product. The short version: an AI e-commerce employee watches the store, tells you what changed and why it matters, and proposes what to do next. The numbers come from the store. The decisions come from you. The problem it exists to solve Running a small or mid-sized online store is a job made of interruptions. The owner or a very small team handles product, marketing, customer service, suppliers, fulfilment and finance, usually at the same time. Shopify provides an enormous amount of data and a capable admin. What it cannot provide is attention. The practical failure is rarely that data is missing. It is that the important signal was there and nobody saw it in time: A best-selling variant sold through faster than usual and went out of stock on a Friday evening. Refunds on one product crept up over two weeks because of a sizing issue nobody connected. A shipping rate change pushed checkout abandonment up, and the store kept paying for ads that sent people to a checkout they left. A batch of orders sat unfulfilled because a supplier tag was missing. An experienced operations person would catch most of these. They would open the admin each morning, look at what moved, compare it with last week, and walk over with a short list: "These three things need you today." Most small stores cannot hire that person. That is the gap an AI e-commerce employee fills. A working definition We define an AI e-commerce employee as software that does four things, in this order: Reads the store's real data — orders, products, variants, inventory, refunds, checkouts, fulfilment status — through the platform's official access. Calculates what changed, using deterministic code: totals, rates, comparisons with previous periods, thresholds crossed. Explains those changes in plain language, ranked by what needs attention, with the numbers shown. Suggests specific next actions, each with its evidence, for a person to approve, adjust or dismiss. What is absent from that list is as important as what is present. There is no step called "decides". There is no step where the model estimates a number the store could have supplied. This follows directly from the principle we apply across the group: AI explains, verified data decides. What it is not Because the phrase is loose, it helps to be explicit. It is not an autopilot. MerchNivo does not reprice products, change inventory, cancel orders, issue refunds or edit customer records by itself. Those actions have direct financial and customer consequences. A person approves them. It is not a chatbot on top of a dashboard. You can ask questions, but the product does not wait to be asked. The core of the service is proactive: it shows up each day with what matters. It is not a source of numbers. A language model is good at explaining and prioritising. It is not a calculator and it is not a database. Every figure in a briefing — revenue, units, refund rate, stock on hand — is retrieved or computed by code from the store's data before the model sees it. If a figure is not available, the briefing says so rather than filling the gap. It is not a replacement for the owner's judgement. A store owner knows things no data feed contains: a supplier is about to change, a product is being discontinued on purpose, a big wholesale order is coming. The software proposes; the person with context disposes. It is not a marketing tool dressed as operations. Growth suggestions are welcome when they follow from the data, but the job is operational first. Our guide to Shopify operations signals describes the signals we treat as core. The daily briefing If MerchNivo had only one feature, it would be the daily briefing. Everything else supports it. A good briefing is short. It fits on a phone screen without much scrolling. It leads with what needs attention, not with a list of metrics. It shows the number, the comparison and the reason it made the list. And it ends with a small number of suggested actions. Here is the shape of a briefing for a hypothetical accessories store, written the way we want MerchNivo to write: Needs attention today Stock risk: Braided USB-C cable, 2 m, black. 14 units left. At the last 7 days' sales pace, that is roughly three days of stock. No incoming stock recorded. Suggested: review reorder with supplier. Refunds up on Laptop Sleeve 14". 6 refunds in the last 7 days against 1 in the previous 7. Four mention "too small" in the refund note. Suggested: check the size guide on the product page. 7 orders unfulfilled for more than 48 hours. All contain the same bundle SKU. Suggested: check whether the bundle is mapped to a fulfilment location. Worth knowing Revenue yesterday was in line with the same weekday last week. Checkout completion was slightly lower than the previous 7-day average; no single cause identified. Everything in that example is a pattern, not a claim about a real store. But notice what the structure does. Each item has a number from the store, a comparison, a reason and a proposed next step. The "worth knowing" section is short and honest about uncertainty. Nothing is invented to fill space. What makes a briefing useful rather than noisy The temptation with any monitoring product is to show everything. That produces a report nobody reads by the second week. We design briefings around a few rules: Rank by consequence, not by size. A small stock-out on a best-seller outranks a large but expected dip after a sale ended. Compare with something meaningful. Same weekday last week, trailing seven days, or the store's own normal range — not an industry benchmark the store has never matched. Show the number every time. "Refunds are up" is not information. "6 refunds in 7 days, up from 1" is. Admit when there is no clear cause. A briefing that says "no single cause identified" earns more trust than one that guesses. Stay quiet on quiet days. If nothing needs attention, the briefing should say so in two lines. Suggested actions and human approval A suggestion is only useful if the person can act on it quickly and trust it. Each suggested action in MerchNivo carries: The action, stated specifically: "Review reorder for SKU X", "Check size guide on product Y", "Pause the discount code Z that expired yesterday but is still active." The evidence: the numbers and records that triggered it. The reason: why this is suggested now. The options: approve, adjust, snooze or dismiss. Approval is not a formality we put up with. It is the core of the design, for three reasons. Consequences are real. A wrong price, a cancelled order or a refund issued in error costs money and customer trust. The cost of asking for one tap of approval is small; the cost of a confident mistake is not. Context lives with people. The store might be deliberately letting a product sell out before a redesign. The model cannot know that unless someone says so. Approval gives the owner a natural moment to apply context. Trust is built, not assumed. When the software is new, people want to see its reasoning. Over time, as they see suggestions that are consistently sound, they approve faster. That trajectory is healthier than asking for full autonomy on day one. When a person dismisses a suggestion, that is useful signal too. A store that repeatedly dismisses low-stock warnings for a particular product is telling us the threshold is wrong for that product, or that the product is being discontinued. The software should adapt its thresholds for that store, not argue. Where the numbers come from This is the part we are strictest about, because it is where AI products most often fail quietly. A language model can produce a sentence like "Revenue was up 12% yesterday" whether or not revenue was up 12%. It will sound equally confident either way. In a store owner's daily briefing, one fabricated number is enough to destroy trust in every other number. So the architecture separates the jobs: Data retrieval pulls records from the store through the platform's official access, within the permissions the merchant granted. Calculation happens in ordinary code: sums, counts, rates, comparisons, thresholds. The output is a structured set of facts. Selection and ranking decide which facts matter today, again mostly with rules the team can inspect. Explanation is where the model comes in. It receives the verified facts and writes the briefing in plain language. Validation checks the model's output against the facts it was given. Any number in the text must match a number in the input. Output that fails is rejected or regenerated. See structured outputs and schema validation. Stock deserves special mention. MerchNivo never guesses inventory. If the store's inventory data is stale, inconsistent across locations or missing, the briefing says so plainly rather than estimating. We explain why in inventory as source of truth. Data boundaries An AI e-commerce employee has access to sensitive business data and customer records. How that access is handled is part of the product, not an afterthought. Each store is a separate tenant. One merchant's data never informs another merchant's briefing, and never appears in another merchant's prompts. See tenant isolation explained. Minimal scopes. We request the platform permissions the features need and no more. Customer data used only for the merchant's own operations. Personal details of a store's customers are not used to train general models and are not shared across the group. Oryvelon's rule of no cross-product user data applies fully. Retrieval stays scoped to the tenant. When the model is given context, that context comes only from the current merchant's store. Nothing retrieved for one store can surface in another store's briefing. MerchNivo runs on the group's shared AI gateway for routing, budgets, safety hooks and observability. The gateway is shared infrastructure; prompts, keys and data remain MerchNivo's own. When the AI is unsure or unavailable A daily briefing is only valuable if it arrives. Model providers have outages, responses can fail validation and costs can spike. We design for all of that. If the model is unavailable, MerchNivo can still send a plainer briefing built directly from the calculated facts: a list of flagged items with their numbers, without the narrative. It is less pleasant to read and still useful. That approach is described in AI fallbacks and degraded modes. If the model's output fails validation — for example a number that does not match the input — we do not send it. A shorter, correct briefing is always better than a fluent, wrong one. And when a question goes beyond what the data can answer, the software says so and points the merchant to the relevant report or to a person. Knowing when to hand over is part of the job; see human escalation in AI products. Who it is for An AI e-commerce employee is most valuable where attention is the scarcest resource: Owner-operated stores where one person is responsible for everything and cannot open five reports every morning. Small teams where operations, marketing and customer service overlap and nobody owns "watching the numbers". Growing stores where order volume has outgrown the owner's ability to spot problems by feel. Agencies and consultants managing several stores who need a quick view of what needs attention in each. It is less useful for large operations with dedicated analysts and custom reporting, which already have people doing this job. For them, MerchNivo may still help as a second pair of eyes, but that is not where we focus. How we test it on a real store Commerce software written only from the outside tends to solve imagined problems. Oryvelon operates Noveniq, a direct-to-consumer technology accessories store, and it gives us a working view of the daily realities MerchNivo is built for: what a stock problem actually looks like at 9 a.m., which alerts are genuinely useful and which ones get ignored. The two companies stay separate. Noveniq's customers are Noveniq's; their data is not pooled into MerchNivo. What crosses over is learning about operations and product design, not customer records. We describe that arrangement in running a real store as a testbed. What a normal week looks like It helps to picture the rhythm rather than the features. Imagine an owner-operated store that connects MerchNivo on a Monday. Monday. The first briefing is cautious. The software has history to compare against but no feedback yet, so it flags a little more than it will later and explains each threshold it used. Tuesday. The owner dismisses a low-stock warning on a product being retired on purpose and marks it as discontinued. That product stops appearing in stock alerts. Wednesday. A refund cluster on one item surfaces with the refund notes attached. The owner edits the product description and approves a suggested follow-up check in a week. Thursday. Nothing needs attention. The briefing is two lines long, which is exactly right. Friday. A weekly view arrives alongside the daily one: slower-moving trends, products whose sales pace has shifted, fulfilment times compared with the previous week. The pattern we aim for is that the owner spends a few minutes a day with the briefing and acts on one or two items, rather than spending an hour looking for them. Questions to ask any AI e-commerce product Whether you are evaluating MerchNivo or something else, these are the questions we would ask: Where do the numbers come from? Are they calculated by code from your store data, or generated by the model? What can it change without you? Get a precise list. "Nothing without approval" is a strong answer. What happens when data is missing or stale? Does it say so, or does it estimate? Is your data kept separate from other merchants' data? And is it used to train anything? What does it do when the AI provider is down? Does anything still arrive? Can you see the evidence behind each suggestion? If not, how will you learn to trust it? How is it priced, and what happens at the limit? See subscription plus usage pricing for how we approach it. Summary An AI e-commerce employee, as we define it at MerchNivo, is software that reads a Shopify store's real data, calculates what changed with ordinary code, explains it in a short daily briefing ranked by consequence, and suggests specific actions with their evidence. A person approves those actions; the software does not change prices, stock, orders or customer records on its own. Numbers always come from the store and are checked against the model's output before anything is sent, stock is never guessed, each merchant's data stays isolated, and when the model is unavailable a plainer briefing still arrives. The goal is not to replace the owner's judgement but to make sure the owner's attention goes to the few things that need it today. ## Shopify operations signals: what to watch daily and weekly URL: https://oryvelon.com/insights/shopify-operations-signals/ Category: Commerce & SaaS Revenue is a lagging number. By the time a store's daily sales visibly drop, the cause usually happened days earlier: a best-seller went out of stock, a shipping change raised checkout abandonment, a batch of orders stalled, a product started drawing refunds. The signals were there. Nobody was looking at them. This article sets out the Shopify operations signals we think every store should watch, split into daily and weekly checks, with notes on how to calculate them and where the thresholds should sit. It is written for store owners and operators, and it is also the thinking behind MerchNivo, the Oryvelon company building an AI e-commerce employee for Shopify stores. Nothing here requires special tools. You can check most of these signals in the Shopify admin and a spreadsheet. The difficulty is doing it consistently, every day, with enough context to know what is normal. That consistency is what software is good at. What counts as an operations signal An operations signal is a measurable change in how the store is running that is likely to cost money, customers or time if nobody acts. It differs from a marketing metric in one important way: it usually points to a specific, fixable thing. "Conversion rate fell" is a marketing observation. "Checkout completion fell after shipping rates changed on Tuesday" is an operations signal. "Sales dropped" is an outcome. "Our best-selling variant has two days of stock left" is a signal. Good signals share four properties: Calculable from store data, without guessing. Comparable with the store's own recent history. Actionable — there is a plausible next step. Timely — acting now is better than acting next week. Daily signals These are the checks that catch problems costing money within days. 1. Stock cover on important variants The single most valuable daily check for most stores. Stock cover is how many days current inventory will last at the recent sales pace: stock cover (days) = units available ÷ average daily units sold (last 7 or 14 days) Three details make it useful rather than misleading: Work at variant level. A product with 200 units across colours can still be out of stock in the colour that sells. Use available, not on-hand, units where the store tracks committed stock, so units already allocated to unfulfilled orders are not counted twice. Account for incoming stock only when it is recorded, with an expected date. A purchase order someone mentioned in a message is not stock. A reasonable starting threshold is to flag any variant whose cover is shorter than the supplier's typical lead time plus a small buffer. For a supplier that takes ten days, a variant with seven days of cover is already late. MerchNivo calculates stock cover from the store's inventory records and never estimates missing counts. If inventory data looks inconsistent — negative stock, locations that disagree, counts that have not changed in suspiciously long — it reports that as its own signal. We explain why in inventory as source of truth. 2. Orders waiting too long to be fulfilled Every store has a normal time from order to fulfilment. Orders that exceed it are an early sign of a problem: a SKU not mapped to a location, a supplier delay, a payment held for review, or simply a busy week. The daily check is: how many unfulfilled orders are older than our usual fulfilment time, and what do they have in common? The second half matters most. Seven delayed orders that all share one SKU point to a specific fix. Seven delayed orders spread randomly point to capacity. 3. New refunds and their reasons Refunds are expensive twice: the lost sale and the handling cost. A daily look at new refunds, grouped by product and reason, catches clusters early. Refund notes and return reasons are some of the most useful text a store has, and they are usually ignored because nobody has time to read them. A useful rule is to flag any product whose refunds in the last seven days are clearly above its own previous seven days, with a minimum count so one refund on a low-volume product does not trigger an alarm. 4. Checkout completion Of the sessions that reach checkout, how many complete an order? A sudden drop usually has an operational cause: a shipping rate change, a payment method failing, a broken discount, a new required field. Abandoned checkouts are a normal part of e-commerce; a change in the rate is the signal. Compare with the same weekday in recent weeks rather than yesterday, because many stores have strong weekday patterns. 5. Payment failures and holds Failed payments, orders flagged for fraud review and payment captures pending for longer than usual can each block revenue quietly. These are low-volume but high-consequence: one stuck high-value order is worth a line in the briefing. 6. Discount codes behaving unexpectedly Discount codes leak. A code meant for a small group gets posted on a coupon site. A code meant to expire keeps working. A stacking rule allows two discounts where the plan was one. The signal is a code whose usage is far above its expected pattern, or a code used after its intended end date. Weekly signals Weekly signals catch slower drifts that daily noise hides. Product-level sales pace Which products are selling faster or slower than their own recent average, over a two-to-four-week window? A product slowly losing pace may need attention on price, photography or stock position long before it becomes a problem. A product gaining pace may need a reorder decision earlier than planned. Slow-moving and dead stock Variants with stock on hand and few or no sales over the last 30 or 60 days tie up cash and storage. The weekly view lists them with their stock value at cost, if the store records costs, so the owner can decide whether to bundle, discount or stop reordering. Refund and return rate by product The daily check catches spikes. The weekly check catches products whose refund rate is persistently higher than the store's normal. Those are usually product problems — sizing, description accuracy, quality — rather than one-off events. Fulfilment time trend Is the average time from order to fulfilment drifting upwards? A gradual increase often signals that the operation is outgrowing its process before anyone feels it. Repeat purchase and customer mix What share of orders came from returning customers this week compared with recent weeks? This is less urgent but useful context: a store whose new-customer share rises while repeat purchases fall may be buying growth through advertising while retention weakens. Discount depth What share of revenue came with a discount, and how large were the discounts? Discount creep is one of the quieter ways margin erodes. Signals at a glance Signal Cadence How it is calculated Typical next action Stock cover Daily Available units ÷ recent daily sales, per variant Review reorder Delayed fulfilment Daily Unfulfilled orders older than usual fulfilment time Check shared SKU, location or supplier Refund cluster Daily Refunds per product vs previous period, with minimum count Read reasons, check listing Checkout completion Daily Completed checkouts ÷ checkouts started, vs same weekday Check shipping, payment, discounts Payment issues Daily Failed, held or pending captures Review specific orders Discount anomalies Daily Code usage vs expected pattern and dates Pause or adjust code Sales pace Weekly Units per product vs own 2–4 week average Adjust stock plan or listing Dead stock Weekly Stock on hand with little or no sales in 30–60 days Bundle, discount or stop reorder Refund rate Weekly Refunds ÷ orders per product over a longer window Fix product or description Fulfilment trend Weekly Average order-to-fulfilment time Review process or capacity Customer mix Weekly Returning vs new customer share Review retention efforts Discount depth Weekly Discounted revenue share and average discount Review promotion policy Setting thresholds that do not cry wolf The fastest way to make a monitoring system useless is to alert on everything. Within two weeks, people stop reading. A few principles keep signals meaningful: Compare with the store's own normal. An industry average abandonment rate says little about a particular store. The store's own trailing average says a lot. Use minimum counts. Percentage changes on small numbers are noise. Two refunds after one is a 100% increase and usually means nothing. Require a minimum absolute count before a signal fires. Match comparison windows to patterns. Weekday patterns, paydays and campaign schedules all shape daily data. Same-weekday comparisons remove a lot of false alarms. Let the owner tune. If a store repeatedly dismisses a signal for a specific product, the threshold for that product is wrong, or the product is being retired. Good software learns this from dismissals rather than repeating the same warning. Separate "needs attention" from "worth knowing". Not every change needs action. Keeping a short, clearly labelled section for context stops the action list from filling up with observations. How MerchNivo turns signals into a briefing The signals above are the raw material. MerchNivo's job is to calculate them reliably, decide which few matter today, and present them in a form a busy owner can act on in minutes. The pipeline is deliberately split so that the model never produces numbers: Retrieve orders, variants, inventory levels, refunds, checkouts and fulfilment records through Shopify's official access, within the permissions the merchant granted. Calculate every signal in code, producing structured facts: variant, metric, value, comparison, threshold. Rank the facts by likely consequence. A stock-out on a top variant beats a small change on a minor product. Explain using the AI model, which turns the ranked facts into a short briefing with suggested actions. Validate the draft against the facts: every number in the text must appear in the input. Anything that does not match is rejected. See structured outputs and schema validation. This follows the principle we apply across Oryvelon: AI explains, verified data decides. The model is good at explaining why three signals might be related, or at turning a list of refund notes into a clear summary. It is not trusted to count. A worked example: a refund spike Imagine a store selling laptop sleeves. On a Wednesday, MerchNivo's refund check finds that one sleeve has six refunds in the last seven days, against one in the previous seven. That clears the minimum-count threshold. The calculation step attaches the refund notes. Four mention the fit. The model reads those notes and writes: Refunds up on Laptop Sleeve 14". 6 refunds in the last 7 days, up from 1. Four refund notes mention the sleeve being too small. Suggested: check the size guide and compatible laptop list on the product page. The owner checks, finds that a recent description edit removed the internal dimensions, restores them and approves a follow-up reminder to re-check refunds in a week. MerchNivo did not edit the product page. It pointed at the right place, with evidence. A worked example: delayed fulfilment Another morning, seven orders are older than the store's usual fulfilment time. The calculation step notices that all seven contain the same bundle SKU. The briefing says so and suggests checking whether the bundle is mapped to a fulfilment location. That is a two-minute fix that would otherwise have been discovered when the first customer emailed to ask where their order was. Signals during campaigns and peak periods Sales events, product launches and seasonal peaks change what normal looks like. A store running a weekend promotion will see sales pace, discount depth and fulfilment backlog all jump at once, and a monitoring system that compares with the previous quiet week will flag everything. Two adjustments help. First, mark campaigns in advance, with dates and the products involved, so comparisons can be read in that light. The briefing can then say "fulfilment backlog is higher than usual, in line with the promotion that started Friday" instead of raising an alarm. Second, tighten the signals that matter more under load. Stock cover on promoted variants deserves a lower threshold during a campaign, because a stock-out in the middle of paid traffic wastes both the stock and the advertising spend. Consistent campaign tagging also makes the weekly review cleaner, since traffic and orders can be tied back to the effort that produced them. We describe the tagging side in UTM standards across brands. Signals and human approval Every suggested action in a MerchNivo briefing waits for a person. The software does not reorder stock, pause discounts, issue refunds or change product pages. It shows the signal, the evidence and a proposed step. This is not caution for its own sake. Operations decisions depend on context the data does not hold. A variant might be running low on purpose because a new version is arriving. A discount code might be spreading because an influencer shared it by arrangement. The owner knows; the software asks. For situations where the right next step is a human conversation rather than an action, see human escalation in AI products. Where the signals get tested We learn which signals matter by running a real store. Oryvelon operates Noveniq, a direct-to-consumer technology accessories brand, and its day-to-day operations are a constant source of practical questions: which alert would have caught this earlier, which one did we ignore, which threshold was wrong. Noveniq's customer data stays with Noveniq. What informs MerchNivo is operational learning — the shape of a problem and the design of a useful alert — not customer records. We describe that boundary in running a real store as a testbed and the wider principle in shared infrastructure, separate data. Common mistakes when monitoring a store Watching revenue first. It is the most visible number and the latest to move. Product-level stock checks. Stock-outs happen at variant level. Ignoring refund notes. They often contain the diagnosis. Comparing with yesterday. Weekday patterns make day-over-day changes misleading. No minimum counts. Small numbers produce dramatic percentages. Too many alerts. A list of twenty items is a list of zero. Counting stock that has not arrived. Incoming stock without a recorded date is a hope, not inventory. Letting software act unsupervised. Automated fixes to operations problems create new ones. A starting checklist If you are setting up operations monitoring for a Shopify store, with or without software, this is a reasonable first version: List your top variants by recent units sold. Record the typical supplier lead time for each. Each morning, check stock cover on those variants against lead time. Check unfulfilled orders older than your usual fulfilment time, and look for a shared cause. Read new refund notes and group them by product. Compare checkout completion with the same weekday last week. Review failed or held payments. Once a week, review sales pace, dead stock, refund rates, fulfilment time and discount depth. Do it by hand for two weeks. You will quickly see which checks matter for your store, which is exactly the knowledge good software should encode. Summary Shopify operations signals are early warnings that appear before revenue moves: stock cover on key variants, delayed fulfilment, refund clusters, checkout completion, payment issues and discount anomalies each day, with sales pace, dead stock, refund rates, fulfilment trends, customer mix and discount depth each week. They are most useful when calculated at the right level, compared with the store's own normal, filtered by minimum counts and ranked by consequence. MerchNivo calculates every signal from store data with code, uses AI only to explain and prioritise, validates every number before sending, and leaves every action to a person's approval — so the owner sees the few things that need attention today, with the evidence to act. ## Running a real store as a testbed for commerce software URL: https://oryvelon.com/insights/real-store-as-testbed/ Category: Commerce & SaaS Commerce software is full of features designed in meeting rooms. They demo well, they fill a pricing page, and they solve problems that store owners rarely have. The problems store owners do have — a variant that sells out at the wrong moment, a refund pattern nobody connected, a bundle that never reaches the warehouse — are messy, specific and easy to miss from the outside. Oryvelon operates Noveniq, a direct-to-consumer brand selling technology accessories. It is a real store with real products, suppliers, orders and customers. It is also, as a secondary benefit, a practical testbed for what we build at MerchNivo, our AI e-commerce employee for Shopify stores. This article explains how running a real store as a testbed works in practice, why it is valuable, where its limits are, and — most importantly — how we get the learning without mixing customer data between companies. Why software teams should run the thing they sell to There is a long tradition of software built by people who use it. It tends to be sharper, because every rough edge costs the builders time personally. The same logic applies to commerce software, with one twist: to feel what a store owner feels, you need a store. Interviews with merchants are essential, and we do them. But interviews tell you what people remember and what they think is important. Operating a store tells you what actually happens on a Tuesday morning when a supplier emails to say a shipment is delayed, three customers ask where their orders are and the ad budget is still spending on a product that just went out of stock. A live store provides three things interviews cannot: Unfiltered problems. You meet the problem before anyone has described it neatly. Real timing. You learn how quickly something needs to be noticed to be useful. A stock alert at 9 a.m. is worth more than the same alert at 6 p.m. Honest usage. When the team running the store ignores an alert, that is a clear verdict on the alert. Noveniq is a business first The most important rule is the one people most often break when they set up a "demo store": the testbed must be a real business with its own goals. A store that exists only to test software drifts. It stops ordering stock when stock is inconvenient. It never has real customer complaints. Its data is too clean and its problems are too convenient. Worse, it tempts the team to judge the software by how well it works on a store designed to make it look good. Noveniq has its own source-of-truth document, its own revenue targets, its own checkpoints and its own customers. It is evaluated at 30, 60 and 90 days on whether to continue, stop or scale, like every other Oryvelon company. If Noveniq were only useful as a testbed and not viable as a business, that would be a reason to rethink it, not to keep it running as a lab. This matters for the testbed value too. The problems worth learning from are the ones a real store faces because it is trying to grow and make money. We describe the wider approach to operating a direct brand in a DTC brand as an operating business. What crosses over, and what does not This is the part we are most careful about. Oryvelon runs on a principle of shared infrastructure, separate data, and we do not allow cross-product user data. Noveniq's customers bought technology accessories from Noveniq. They did not agree to become part of a software company's dataset. So the boundary is explicit: Crosses from Noveniq to MerchNivo Stays with Noveniq Descriptions of operational problems ("a bundle SKU was not mapped to a location and orders stalled") Customer names, addresses, emails, phone numbers Feedback on alert usefulness ("this refund alert came too late") Order histories linked to individuals Timing and workflow observations Payment details Requests for features and changes Support conversations and refund notes that identify people Synthetic or anonymised examples written for testing Marketing lists and consent records General patterns described in words Raw data exports The pattern is simple to state: learning crosses, records do not. A Noveniq operator might tell the MerchNivo team "we missed a refund cluster because the notes were buried; the briefing should summarise refund reasons". That is valuable, and it contains no customer data. What the operator does not do is send the refund notes themselves. How MerchNivo runs on Noveniq If Noveniq chooses to use MerchNivo as a customer, it connects in exactly the way any other merchant would: through Shopify's standard access, granted by the store, into its own isolated tenant. MerchNivo processes that data to serve Noveniq, the same way it would serve any store, and Noveniq's data is subject to the same tenant isolation and retention rules as everyone else's. There is no back door, no special internal access and no bulk export into a shared analytics store. MerchNivo engineers do not get broader access to Noveniq's data because both companies are in the same group. Access follows the same least-privilege rules that apply everywhere. Test data for development When MerchNivo needs realistic data to develop and test features, it uses synthetic data shaped like a store's data — plausible products, variants, order patterns and edge cases — not copies of Noveniq's records. Noveniq's experience helps make that synthetic data more realistic ("stores like ours have bundles, multi-location stock and seasonal spikes"), but the data itself is generated. We cover this in environments and test data. What a testbed is good at Used well, a real store is extremely good at a specific set of things. Finding the problems worth solving. Many of the operations signals MerchNivo watches are the kind of thing an operator discovers only by running a store: the order stuck because of one mapping issue, the discount code that kept working after a campaign, the variant that sold out in one colour while the product looked healthy overall. Testing whether an alert is useful. A signal can be accurate and still useless. A real operator tells you quickly: too late, too noisy, not specific enough, or exactly right. Setting sensible defaults. What threshold for stock cover feels right? How many refunds before a cluster is worth mentioning? A live store gives a first answer that is grounded in something real, before it is checked more widely. Catching awkward edge cases. Real stores have messy data: products renamed mid-season, variants archived and restored, locations added and removed. Software that works on clean demo data often breaks on these. Rehearsing failure. What does the operator see when the AI model is unavailable, or when inventory data looks inconsistent? Watching a real person react to a degraded briefing is more instructive than any test script. What a testbed cannot prove It is equally important to be honest about the limits. One store is one data point. It is not a market. Noveniq sells technology accessories direct to consumers. Its product mix, order volume, seasonality and team size are its own. A fashion brand, a food brand or a high-volume wholesale operation will have different problems. Anything learned from Noveniq is a hypothesis about other stores, not a conclusion. It is biased towards the builders. The people running Noveniq know the MerchNivo team and understand the product's intent. They are more forgiving and more articulate than a typical merchant. Features that make sense to them may confuse someone who has never heard of the product. It cannot validate pricing. Whether merchants will pay, and how much, can only be learned from merchants who are paying. A store the group owns tells us nothing about willingness to pay. It can create false confidence. If every feature is shaped around one store, the product ends up fitting that store and nothing else. So we treat Noveniq's feedback as a strong early signal and then check it. Before a pattern from Noveniq becomes a default in MerchNivo, we look for the same problem among other merchants through interviews, support conversations and usage patterns within each merchant's own tenant. If nobody else has the problem, it probably does not belong in the product. How the feedback loop works In practice, the loop is informal but disciplined. It runs on the same operating cadence as the rest of the group. Noveniq runs its store and notices friction: something missed, something noisy, something slow. The operator writes it down in plain language, without attaching customer records. A short note: what happened, when it would have been useful to know, what they did about it. MerchNivo reviews the notes and decides whether it looks like a general problem, a Noveniq-specific quirk, or a data quality issue. If it looks general, MerchNivo checks with other merchants before building. If it gets built, Noveniq is often among the first to use it, as an ordinary customer in its own tenant. The result is reviewed: did the change help? Did it help anyone else? A worked example makes the rhythm concrete. Imagine Noveniq runs a weekend promotion on phone cases. On Monday, the operator realises that one popular case sold out on Saturday afternoon while ads kept running until Sunday night. Their note to MerchNivo reads, roughly: "Stock-out on promoted variant was not visible until Monday. During campaigns, stock alerts need to be tighter and faster." There is no customer data in that sentence. MerchNivo asks a few other merchants whether they have had the same experience, finds that several have, and adds a campaign-aware stock threshold. Noveniq then uses the new behaviour during its next promotion, just as any other merchant would. What the store gets in return The arrangement is not one-way. A store that sits next to a software team gets something too, and it is worth being clear about what that is, because it should never distort how the store is run. Noveniq benefits from a team that thinks carefully about operations signals, thresholds and daily routines. The act of writing feedback notes forces the store to reflect on its own processes: why did that bundle fail to map, why did nobody read the refund notes, why was the reorder late. Many of those reflections lead to fixes inside Noveniq that have nothing to do with software at all — a clearer supplier calendar, a better product description, a simple weekly check. It also gets early access to changes, as any design partner would. But it pays for and uses the software on the same terms as other merchants, and its operators are free to say a feature is not worth using. A testbed that is obliged to praise the product is worse than no testbed. Failure modes of internal testbeds We have thought hard about how arrangements like this go wrong, because the failure modes are predictable. The store becomes a showroom. Its data gets cleaned up so the software looks good in demos. Stock counts are corrected by hand before a presentation. Real problems are hidden. The fix is to never use the testbed as a sales demo; demos use synthetic data. Data quietly leaks across. Someone exports "just a sample" of orders to debug a feature. Then a second sample. Then a shared folder of samples. The fix is structural: no access path from MerchNivo into Noveniq's data outside the normal merchant connection, and a written rule that debugging with real data happens only inside Noveniq's own tenant, by people Noveniq authorises, for Noveniq's benefit. The store's priorities bend. Noveniq starts stocking products or running campaigns because they would make interesting test cases rather than because they make business sense. The fix is Noveniq's own source of truth and checkpoints, which judge the store only on its performance as a store. One voice dominates the roadmap. The internal store is always available and always articulate, so its requests get built first. The fix is the confirmation step: a request from Noveniq carries no more weight than a request from any other merchant until others share the problem. Customer communications drift. Noveniq's customers shop at a technology accessories store. Their experience, communications and privacy terms are Noveniq's. If we ever found ourselves wanting to change what Noveniq tells its customers because of MerchNivo, that would be a sign the boundary had slipped. Learning stays in people's heads. The operator mentions a problem in passing, the MerchNivo team nods, and nothing is written down. Three months later nobody remembers why a threshold was set the way it was. The fix is boring and effective: short written notes, kept with the rest of MerchNivo's product records, so every default can be traced back to the problem that prompted it and the merchants who confirmed it. Why keep them as separate companies A reasonable question is why Noveniq and MerchNivo are separate companies at all. Why not make the store part of the software business? The answer comes from how Oryvelon is designed. Every company has its own brand, customers, revenue model and data, and every company should be able to stand alone. Merging a consumer brand into a software company would mix two very different businesses with different customers, different risks and different economics. It would also make the data boundary much harder to hold. Keeping them separate brings clear advantages: Noveniq's customers deal with a store, not a software experiment. Their data is governed by Noveniq's own policies. MerchNivo's customers can trust that it has no special relationship with any one merchant's data, including one owned by the same group. Each business is judged on its own results. A strong store does not hide a weak product, or the other way round. Either company could be sold, spun out or closed without untangling the other. This is the same reasoning behind why we don't merge products into one app, applied to a store and a software product. Lessons for other founders If you build software for merchants and are considering running a store to learn from, a few suggestions from how we approach it: Run it as a real business. Give it real goals and real checkpoints. A fake store teaches fake lessons. Write the data boundary down before launch. Decide what crosses and what does not, and put it in both companies' source-of-truth documents. Connect through the front door. If the store uses your software, it should connect exactly like any other customer. Use synthetic data for development, shaped by what you learned, not copied from the store. Treat findings as hypotheses. One store suggests; many merchants confirm. Watch for bias. Your own team will forgive things a stranger will not. Keep the store's customer communications its own. Never use its customer list to market the software. Summary Running a real store as a testbed gives commerce software something interviews and demos cannot: unfiltered problems, real timing and honest verdicts on what is useful. Oryvelon operates Noveniq as a genuine direct-to-consumer business first, and its daily experience informs what MerchNivo builds. What crosses between them is operational learning and product feedback written in plain language, never customer records; if Noveniq uses MerchNivo, it connects like any other merchant into its own isolated tenant, and development relies on synthetic data. Because one store is only one data point, findings from Noveniq are treated as hypotheses and checked with other merchants before they shape the product. ## Subscription plus usage pricing for AI SaaS URL: https://oryvelon.com/insights/saas-subscription-plus-usage/ Category: Commerce & SaaS Traditional software had a convenient property: once it was built, serving one more customer cost almost nothing. A flat monthly subscription worked because the heaviest user and the lightest user cost roughly the same to serve. AI products break that assumption. Every briefing generated, every report written and every question answered calls a model, and every call has a cost. A customer who uses the product ten times as much costs roughly ten times as much to serve. A flat price that is comfortable for a typical customer can become a loss on a heavy one. That is why many AI software products, including several in the Oryvelon group, use subscription plus usage pricing: a recurring fee with an included allowance, and a clear way to pay for more. This article explains how we think about it — what the subscription should cover, how to pick a usage unit, how to set fair limits, and the mistakes that make usage pricing feel like a trap. We do not publish specific prices here. Pricing belongs to each company and changes with its market. The principles are what carry across. Why flat pricing breaks for AI products Consider a simplified example. Imagine an AI product priced at a flat monthly fee. Most customers use it moderately and cost a comfortable fraction of that fee in model calls, hosting and support. A small group of customers use it constantly. Their model costs alone exceed what they pay. Three things tend to happen: Margins depend on light users subsidising heavy ones. That works until the mix shifts, which it often does as the product improves and people use it more. The company quietly degrades service — cheaper models, shorter outputs, hidden throttling — to protect margins. Customers notice and trust erodes. Pricing becomes a guessing game in which every new feature that calls a model makes the flat price less sustainable. None of this means flat pricing is always wrong. For a product with low, predictable AI usage per customer, a flat price with a sensible fair-use policy is often best. The point is that the choice should be made knowing the cost structure, not by copying traditional SaaS. We cover the cost side in cost discipline for AI products. The structure: base, allowance, beyond Subscription plus usage has three parts. The base subscription pays for access to the service and the value that does not scale with use: the product itself, integrations, data storage, the dashboard, support, security and ongoing improvement. The included allowance is a quantity of the usage unit bundled into each plan. It should be generous enough that most customers on that plan never think about it. Beyond the allowance, customers have a clear option: upgrade to the next plan, buy a top-up, or pay a published per-unit rate. Which option fits depends on the product. Plans usually step up in both base price and allowance. The goal is that each customer lands on the plan that matches their normal usage, pays a predictable amount most months, and has a fair path when a month is unusual. Choosing the right usage unit The most important decision in usage pricing is the unit. Get it right and customers understand their bill without thinking. Get it wrong and every invoice is a surprise. A good usage unit is: Meaningful to the customer. It corresponds to something they value and can see. Predictable. Customers can estimate how many they will use. Roughly proportional to cost, so the company is not exposed to heavy losses on one type of use. Easy to count and display in the product itself. Tokens fail most of these tests. They are an internal measure of model input and output. Customers cannot see them, cannot predict them and should not have to learn what they are. A bill that says "4.2 million tokens" tells a store owner nothing. Better units usually map to the product's core output. Across Oryvelon's companies, the natural units differ: Company Core value Natural usage unit MerchNivo Daily operational help for a Shopify store Connected stores, plus briefings, analyses or questions within an allowance ZodiVela Personal discovery readings Credits per reading, or readings included in a subscription KeşifAtlası Visa and relocation eligibility Free test, then a paid report per case EduRelia AI learning for students and schools Seats or learners per school licence Each unit is something the customer can picture. A store owner knows how many stores they run. A ZodiVela user understands that a reading costs a credit. A family understands that a KeşifAtlası report covers one relocation case. The internal cost of each unit is modelled by the company; the customer sees the unit. Sizing the allowance The included allowance is where fairness is won or lost. Start from real or carefully estimated usage. For each plan, look at what a typical customer on that plan uses in a normal month, and set the allowance comfortably above it. The aim is that the large majority of customers never reach the allowance in an ordinary month. Then check the economics at the allowance, not just at typical use. If a customer uses exactly their full allowance, does the plan still cover its costs with a reasonable margin? If not, either the allowance is too high or the base price is too low. This calculation belongs in each product's unit economics. Finally, leave room for model costs to change. Model prices move, sometimes downwards, sometimes not. Routing some work to cheaper models, caching repeated work and trimming prompts can all lower cost per unit over time. We describe the first of those levers in model routing. When costs fall, a good company can raise allowances rather than cut prices, which customers generally appreciate more. Fair limits, not traps The difference between usage pricing that customers accept and usage pricing they resent is almost entirely about how limits behave. We hold ourselves to a few rules. Limits are visible. Customers can see their current usage and remaining allowance at any time, in the product, in the unit they are billed in. Warnings arrive early. A notice at a sensible point before the allowance runs out, and another close to it, with the options laid out. Nobody should discover a limit by hitting it. No silent overages. If going beyond the allowance costs money, the customer chooses that — by setting a spending cap, approving a top-up or upgrading. A bill should never contain a charge the customer did not see coming. The core service keeps working. When an allowance runs out, the product degrades gracefully rather than switching off. For MerchNivo, that might mean the daily briefing continues in a simpler form built directly from calculated store data, while additional on-demand analyses wait for a top-up or the next cycle. The owner still sees what needs attention today. See AI fallbacks and degraded modes. Spending caps are available. Customers who want certainty can set a maximum monthly spend and know it will not be exceeded. Limits reset predictably, on a stated date, with no confusing rolling windows. Budgets on our side too Usage pricing protects the customer from surprise bills. Budgets protect the company from surprise costs. Every Oryvelon product that calls models does so through the group's shared AI gateway, which enforces budgets per product and, where needed, per customer. If a bug or a misuse pattern starts generating unexpected calls — a loop that regenerates the same briefing, a script hammering an endpoint — the budget catches it before the invoice does. This matters for pricing because a usage allowance only makes sense if the company knows its cost per unit. Observability at the gateway gives each company that number, broken down by feature and model, so plans can be designed from facts rather than guesses. Choosing between upgrade, top-up and pay-as-you-go When a customer exceeds their allowance, there are three common paths. Each suits different products. Upgrade to the next plan. Best when exceeding the allowance signals that the customer's normal usage has grown. A store that has added a second shop or doubled its order volume probably belongs on a higher plan permanently. Top-up packs. Best when usage is lumpy. A ZodiVela user might want extra readings around a particular moment without changing their subscription. Credits bought in advance are clear and self-limiting. Metered overage at a published rate. Best for business customers who prefer continuity over caps, and who have agreed to it with a cap in place. It needs the clearest communication of all. Many products offer more than one. The important thing is that the customer chooses, with the cost visible, before it is incurred. Worked example: an AI briefing product Imagine a hypothetical AI product that sends store owners a daily briefing and answers operational questions on demand. Here is how we would reason about its pricing, without real numbers. Identify the cost drivers. The daily briefing runs once per store per day, a predictable cost. On-demand questions vary widely between customers. Deeper analyses — a monthly product review, say — cost more per request. Separate predictable from variable. The daily briefing is part of the base subscription for each connected store. It is the core value and its cost is stable. Choose the usage unit for the variable part. On-demand questions and analyses are counted as a single understandable unit, with deeper analyses counting as more than one, clearly labelled. Size allowances per plan so that a typical store on each plan stays within them. Design the limit behaviour. When the allowance is used, the daily briefing continues; on-demand requests show remaining allowance, then offer a top-up or upgrade. Check the worst case. A store at full allowance every month should still be a profitable customer on its plan. The result is a price the customer can predict, with the variable cost controlled and the core service protected. Consumer and business customers need different shapes The same model does not fit every audience. Consumers and businesses think about spending differently, and the pricing shape should respect that. Consumers want certainty and small decisions. A person using ZodiVela for a reflective reading does not want to monitor a meter. Credits bought in advance, or a subscription with a clear number of included readings, give them a fixed, visible cost. Metered overage is almost always wrong for consumers: it turns a light, personal product into something that feels like a utility bill. Consumer products should also make it easy to stop, pause or cancel, because trust at the exit is part of trust at the start. Businesses want predictability and room to grow. A store owner using MerchNivo is making an operating decision. They care that the cost is stable month to month, that it scales sensibly as the business grows, and that nothing unexpected appears on the invoice. Plans tied to connected stores, with generous allowances and optional caps, fit that mindset. Institutions want simple annual terms. For school licensing at EduRelia, usage is usually expressed as seats or learners for a period, agreed in advance. Schools budget annually; a usage meter would make purchasing harder rather than fairer. The variable AI cost still exists, so it is modelled inside the per-seat price with fair-use protections rather than exposed as a meter. We cover that model in B2B2C licensing for schools. What to measure after launch A pricing model is a hypothesis until customers use it. After launch, a handful of measures tell us whether it is working: Share of customers reaching their allowance each month, per plan. If it is high, the allowance is too low or customers are on the wrong plan. Cost per usage unit, tracked over time and by feature, so we notice when a new feature makes a unit more expensive. Gross margin per plan at typical and full usage, to confirm that no plan loses money on legitimate heavy users. Upgrade and top-up behaviour after limit warnings: do customers upgrade, buy more, or leave? Support contacts about billing, which are the clearest signal that something is confusing. Cancellations that mention price, read individually rather than counted only. These numbers feed each company's regular review. They let us adjust allowances, add a plan or change the unit before a pricing problem becomes a retention problem. Communicating price honestly Usage pricing has a reputation problem, mostly earned by products that hide the details. Clear communication fixes most of it: Show the allowance and unit on the pricing page, in plain language. Give examples: "a typical store with daily briefings and occasional questions fits comfortably in this plan." Explain what happens at the limit before the customer signs up. Keep the invoice readable, with usage in the same unit as the pricing page. Announce changes in advance, and honour existing terms until the next renewal. For consumer products especially, trust at the moment of payment carries over to trust in the product. We discuss the free-to-paid step in from free test to paid report and the recurring side in recurring revenue loops. Common mistakes Pricing by tokens. Accurate for the company, meaningless to the customer. Allowances set at average usage. Half of customers would hit the limit every month. Hard shutdowns at the limit. The product disappears exactly when the customer is using it most. Silent overages. The fastest way to lose a customer and earn a dispute. Degrading quality instead of pricing honestly. Quietly swapping to a weaker model to protect margin erodes trust faster than a clear price change. Ignoring the worst case. A plan that loses money on its heaviest legitimate users is not a plan. Too many units. One or two usage measures per product; more than that and nobody can estimate a bill. How this fits the group Each Oryvelon company owns its pricing. There is no group-wide price list and no bundle that ties one company's plan to another's, because every company needs to stand alone and customers of one should never be pushed towards another. What the group shares is the method: understand cost per unit, choose a customer-meaningful unit, set generous allowances, keep limits visible and protect the core service. Pricing decisions are revisited at each company's continue, stop or scale checkpoints, alongside unit economics, so a plan that looked right at launch is checked against what customers actually do. Summary AI software products carry a real variable cost per action, so a flat subscription alone can make heavy users unprofitable or push companies into quietly degrading service. Subscription plus usage pricing solves this with a base fee for the service and its stable value, an included allowance sized so most customers never reach it, and a clear path beyond it through upgrades, top-ups or capped overage. The usage unit should be something customers understand — stores, briefings, readings, reports or seats — never tokens. Limits must be visible, warned early and never produce silent charges, and when an allowance runs out the core service should keep working in a simpler form. On our side, gateway budgets and per-feature cost data make the economics knowable, so plans are built from facts and revisited at every checkpoint. ## Inventory as source of truth: why AI must never guess stock URL: https://oryvelon.com/insights/inventory-truth/ Category: Commerce & SaaS Some numbers in a business can be approximately right and still useful. Website traffic, estimated reach, a rough forecast of next month's sales — nobody is harmed if they are a little off. Inventory is not one of those numbers. If a store believes it has twelve units and actually has zero, it will sell items it cannot ship. If it believes it has zero and actually has forty, it will hide products customers want and perhaps reorder stock it does not need. Every error turns into a concrete cost: a cancelled order, an apology email, a refund, a wasted purchase, a customer who does not come back. That is why we treat inventory as a source of truth — one authoritative record, kept accurate by disciplined syncing and regular reconciliation — and why MerchNivo, our AI e-commerce employee for Shopify stores, is built so that its AI never guesses stock. This article explains both halves: how to keep inventory true, and how AI should behave around it. Why inventory is special Most e-commerce data describes the past. Orders happened. Refunds happened. Sessions happened. Inventory is different: it is a claim about the present, used to make promises about the future. When a product page says "in stock", the store is promising it can ship. Inventory is also unusually easy to get wrong, because many things change it: sales on every channel the store uses; returns and restocks; damaged, lost or quarantined units; transfers between locations; incoming purchase orders arriving early, late, short or over; manual corrections after a count; bundles and kits that consume several components; integrations with warehouses, marketplaces and point-of-sale systems, each with its own timing. Each of these is a chance for the count to drift. And unlike most data problems, stock drift is invisible until it causes a failure. One system of record The first principle is simple to state and surprisingly often ignored: for each item at each location, exactly one system holds the authoritative count. For many small and mid-sized stores, that system is the store platform itself — Shopify's inventory at each location. For others, it is a warehouse management system or an inventory tool, with the store platform receiving counts from it. Either arrangement can work. What does not work is two systems that each believe they are in charge. Once the system of record is chosen, every other system has a defined relationship with it: System Relationship to the system of record Storefront and checkout Reads available stock; decrements through orders Marketplaces and other channels Receive stock updates; send orders back Warehouse or 3PL Either is the system of record, or reports receipts and adjustments to it Purchasing and supplier tools Record incoming stock; update counts on receipt Reporting and analytics Read only AI tools, including MerchNivo Read only, via the platform's official access Writing this table down, per store, is one of the most useful half-hours an operator can spend. It belongs in the store's source-of-truth document, next to the other rules that should not drift. On hand, committed, available, incoming Stock is not one number. At minimum, a useful model separates: On hand — units physically at a location. Committed — units allocated to orders not yet fulfilled. Available — units that can still be sold (on hand minus committed, minus anything reserved or unsellable). Incoming — units on a recorded purchase order or transfer, with an expected date. Confusing these is a common source of error. A report that shows on-hand units as sellable will overstate stock during busy periods, when many units are already committed to orders waiting to ship. MerchNivo works with the platform's distinctions rather than inventing its own, and labels which figure it is using. Sync keeps systems close When more than one system touches inventory, they need to stay in step. Sync is the ongoing process of sending changes from one to another. Good sync has a few properties: Direction is defined. Each field flows one way, from the system of record outward, or from a defined source inward. Two-way sync of the same field between two systems is where overwrites and loops come from. Changes are events, not snapshots. Sending "sold 2 units" is safer than sending "stock is now 10", because the second can overwrite a change that happened in between. Failures are visible. A sync that silently fails is worse than no sync, because everyone believes the numbers are current. Timing is known. If a channel updates every fifteen minutes, everyone should know that a fast-selling item can oversell within that window, and set buffers accordingly. Updates are idempotent. The same change delivered twice should not decrement stock twice. Even with all of that, sync is not enough on its own. Messages get lost, integrations get misconfigured, someone edits a count directly in the wrong system, a bundle mapping breaks. Sync keeps systems close. It does not prove they agree. Reconciliation proves they agree Reconciliation is the deliberate, regular comparison of inventory records to find differences that sync missed. It happens at two levels. System-to-system reconciliation compares counts between the system of record and each connected system. If the store platform says 18 and the marketplace listing says 25, something has drifted. Ideally this runs automatically and reports differences above a small tolerance. Record-to-physical reconciliation compares the system of record with what is actually on the shelf. That means counting: full stocktakes periodically, and cycle counts — a rotating subset of items — more often. Fast-moving and high-value items deserve more frequent counts. When a difference is found, the resolution should be recorded: what the difference was, what caused it if known, and what adjustment was made. Over time, those records show patterns. A particular location that always runs short, a bundle that always drifts, an integration that loses updates at weekends. Patterns point to fixes. Why AI must never guess stock This is where AI enters, and where our position is firm. A language model is designed to produce plausible text. Ask it how many units of a product are left and, if it does not have the number, it may produce one anyway — in the same confident tone it uses for everything else. It might extrapolate from yesterday's figure, average across variants, or simply fill the gap. The result reads like a fact. In most domains, that is an inconvenience. In inventory, it is an operational failure. A briefing that says "you have plenty of stock of the black cable" when the store has none will lead the owner to keep advertising a product that cannot ship. So the rule across Oryvelon — AI explains, verified data decides — applies with extra force here: Stock figures are retrieved by code from the system of record, through the platform's official access, at a known time. Calculations are done by code: available units, stock cover, days until stock-out, variance from the last check. The model receives those facts and explains them. It does not receive a request to "estimate" anything. Output is validated: every stock figure in the briefing must match a figure in the input. See structured outputs and schema validation. Missing data is stated, not filled. If a figure is unavailable, the briefing says so. What MerchNivo says when data is unclear The most important behaviour is what happens when inventory data looks wrong. MerchNivo checks for signs that stock data cannot be trusted and reports them as their own signal, rather than silently working with a bad number: Negative available stock, which usually means orders were accepted beyond the count. Counts that have not changed in an unusually long time on items that are selling. Disagreement between locations and the total, or between channels. Sales recorded against variants with no inventory tracking. Recent large manual adjustments without an obvious reason. A briefing in that situation reads something like: Inventory data needs checking: Wireless Charger Pad, white. The store shows –3 available units, which suggests orders were accepted beyond recorded stock. Stock cover cannot be calculated reliably until the count is corrected. Suggested: confirm the physical count and update inventory. That is less satisfying than a neat stock-cover number. It is also correct, and it points the owner at the real problem. Worked example: overselling during a promotion Imagine a store running a weekend promotion on a popular variant. The store platform is the system of record. A marketplace channel receives stock updates on a fixed interval. On Saturday, the variant sells quickly on both channels. Between updates, the marketplace keeps selling units the store has already sold through its own checkout. By Saturday evening, available stock on the platform shows –4. What should happen: Monday's briefing (or an alert sooner, if configured) flags negative available stock on that variant as a data problem, lists the orders affected and suggests confirming the physical count. The owner checks the shelf, finds that a few units were in a returns bin and not counted, adjusts stock and decides how to handle the remaining oversold orders. The weekly review notes the cause: sync interval too long for a promoted fast mover. The fix is operational, not AI: a stock buffer on the marketplace for promoted items, or a shorter sync interval during campaigns. At no point did the AI estimate how many units were "probably" on the shelf. It pointed at the discrepancy, with evidence, and a person resolved it. Forecasts are not stock There is one place where estimation is legitimate: looking ahead. How many days will current stock last? When should a reorder be placed? These are forecasts, and forecasts are useful. The rule is that a forecast must be built on true stock and must be labelled as a forecast. MerchNivo keeps the two clearly separate. The figure "14 units available" comes from the system of record and is presented as a fact, with the time it was read. The figure "about three days of cover at the last seven days' pace" is a calculation, done in code, with its method stated in the same sentence. The owner can see which part is observed and which part is projected, and can disagree with the projection — a promotion starting tomorrow will change the pace — without doubting the count. What the AI model contributes is the explanation around those numbers: why a variant's pace has changed, which items share a supplier with a long lead time, which reorder decisions are most urgent. It does not adjust the forecast by intuition, and it never turns a projection back into a stock figure. Multi-location stores Stores that hold stock in more than one place — a home office, a warehouse, a fulfilment partner, a retail counter — multiply every problem above. The total can look healthy while the location that serves most orders is empty. A few practices help. Report availability per location as well as in total. Make fulfilment rules explicit, so it is clear which location ships which orders. Treat transfers as stock in transit, visible but not available, until they are received. And reconcile each location on its own schedule, since the causes of drift are usually local. MerchNivo shows per-location figures when a stock signal depends on them, so an owner is not reassured by a total that hides an empty shelf. Buffers and safety stock Because sync is never instant and counts are never perfect, many stores hold back a small buffer on some channels — publishing slightly less than the true available stock — so that timing gaps do not cause overselling. Separately, safety stock is the level at which a reorder should happen early enough to cover supplier lead time plus uncertainty. Both are business decisions, and both belong to the owner. MerchNivo can show the data behind them — how often a variant has oversold, how long reorders take to arrive, how variable daily sales are — and can suggest reviewing a buffer. It does not set them. Bundles, kits and variants Bundles deserve their own warning, because they are a frequent source of inventory errors. A bundle that sells as one product might consume three components. If the bundle has its own stock count that is not linked to its components, both can drift: the bundle keeps selling while a component runs out, or the component shows stock that is really committed to bundles. The principle is the same as everywhere else: one source of truth. Either the components are authoritative and the bundle's availability is calculated from them, or the bundle is stocked as its own physical item. Mixing the two is where problems start. MerchNivo flags bundles whose orders are stalling in fulfilment, which is often the first visible symptom — see our guide to Shopify operations signals. How we learn from a real store Inventory problems look different on paper than on a Saturday afternoon. Oryvelon operates Noveniq, a direct-to-consumer technology accessories store, and its operations are a steady source of realistic inventory questions: how often cycle counts are actually done, which discrepancies are most common, how operators react to a data-quality warning compared with a neat number. That learning shapes MerchNivo's defaults. Noveniq's customer data does not. The boundary is described in running a real store as a testbed, and development uses synthetic stock scenarios rather than copies of real records, as covered in environments and test data. A checklist for inventory truth For any store, whether or not it uses AI tools: Name the system of record for stock at each location, and write it down. Map every system that reads or writes inventory, with the direction of each flow. Separate on hand, committed, available and incoming in reports and decisions. Check sync failures, not just sync settings. Reconcile systems automatically and report differences above a tolerance. Cycle count fast movers and high-value items more often than the rest. Record every adjustment with a reason. Set channel buffers for fast-selling items where sync has a delay. Link bundles to components, or stock them as their own items — not both. Make sure any AI tool reads stock from the system of record and says clearly when it cannot. Common mistakes Two systems in charge. Each overwrites the other; nobody knows which is right. Snapshot sync. "Stock is now X" overwrites changes made in between. Silent sync failures. Numbers look current and are not. Treating on hand as available during busy periods. Counting stock that has not arrived. An expected delivery is not inventory until it is received. Unlinked bundles. A reliable source of overselling. AI that estimates. A plausible wrong number is worse than an honest "unknown". Fixing symptoms, not causes. Adjusting a count without recording why guarantees it will drift again. Summary Inventory is a claim about the present that becomes a promise to customers, so errors cost money immediately. Treating inventory as a source of truth means naming one authoritative system for each item and location, defining how every other system reads from or reports to it, separating on hand, committed, available and incoming stock, and pairing sync with regular reconciliation so drift is found and explained. AI has a clear place in this: it explains counts that code has retrieved and checked, flags data that looks unreliable and suggests actions a person approves. It never estimates stock. When MerchNivo cannot trust an inventory figure, it says so plainly, because an honest warning protects a store far better than a confident guess. ## Running a DTC brand as a real operating business URL: https://oryvelon.com/insights/dtc-brand-as-operating-business/ Category: Commerce & SaaS Plenty of direct-to-consumer brands look healthy from the outside. The site is clean, the ads are running, orders arrive every day. Then someone adds up landed cost, shipping, payment fees, returns and the discount codes that went out last month, and the picture changes. The brand was busy. It was not profitable. Running a DTC brand as an operating business means refusing to confuse those two things. Noveniq sells technology accessories direct to customers, and inside Oryvelon it is treated as a company in its own right: its own source of truth, its own numbers, its own continue/stop/scale checkpoints. It also happens to be the group's commerce laboratory, which we cover in Why we run a real store as a testbed. This article is about the first job. The store has to work as a business before it is useful as anything else. What makes DTC different from "selling online" Selling online can mean many things: listing on a marketplace, dropshipping someone else's catalogue, running a storefront as a side channel for a wholesale business. DTC is narrower. The brand owns the product, the storefront, the customer relationship and the consequences. That ownership is the appeal. You set prices, you keep the customer data (within the consent the customer gave), you decide what the unboxing feels like. It is also the burden. There is no marketplace absorbing chargebacks, no wholesale buyer taking stock risk, no platform sending you free traffic. Every cost that someone else would normally carry lands on you. So the operating question is simple to state: after an order is paid, picked, shipped, possibly returned and supported, what is left? Everything else in this article is an attempt to answer that question honestly and to make the answer bigger over time. Start with the product, not the ad account Many DTC brands begin with a creative idea and a media plan. We start with a product sheet. For each candidate product in Noveniq's range, before anything goes live, we want to know: Landed cost. Purchase price plus freight, duties, packaging and any inbound handling, per unit. Selling price range. What comparable products actually sell for, not what we would like to charge. Size and weight. These drive shipping cost and, for accessories, often decide whether a product is viable at all. Failure and return risk. Cables, cases and chargers have different defect profiles. A product that breaks or does not fit gets returned. Compatibility complexity. A phone case that exists in twenty model variants is a stock problem, not a product. Repeat potential. Does owning it create a reason to buy again, or a reason to buy something next to it? Technology accessories are a useful category for learning because these factors vary so much. A small, light, universal item with low return risk can carry a modest price and still leave a healthy margin. A bulky, model-specific item with a high defect rate can look attractive on the price tag and still lose money once returns and shipping are counted. We would rather launch a narrower range that works than a broad range that needs constant rescue. Contribution margin is the number that matters Revenue tells you how much money passed through. Gross margin tells you what the product costs relative to its price. Neither tells you whether an order was worth taking. The number we run Noveniq on is contribution margin per order: what remains after every variable cost the order caused. A simplified, hypothetical order might look like this: Line Example amount Order value (after discount) 100 Product landed cost −38 Payment processing −3 Packaging −2 Outbound shipping (net of any charged to customer) −7 Expected returns and refunds cost (averaged) −5 Contribution before marketing 45 Marketing cost to acquire this order −30 Contribution after marketing 15 These figures are illustrative, not Noveniq's actual numbers. The structure is the point. A healthy return on ad spend can still sit on top of a thin contribution if shipping, discounts or returns creep up. And a first order that loses money can still be a good order if the customer is likely to come back. We break this down further, product by product and channel by channel, in Unit economics per product. For a store, the discipline is to calculate it before scaling spend, and to recalculate whenever a cost input changes. Stock is cash sitting on a shelf Inventory is where DTC brands most often get into trouble quietly. Buy too little and you stock out of your best seller just as an ad starts working. Buy too much and your cash is tied up in cases for phones people stopped buying. For Noveniq we keep three rules: One system is the source of stock truth. Numbers in spreadsheets, supplier emails and ad dashboards are copies. When they disagree, the store's inventory record wins, and discrepancies are corrected there. We explain why in Inventory truth. Reorder points are written down per product. They account for supplier lead time and a realistic sales rate, not the best week on record. Slow stock is reviewed monthly. Items that have not moved are marked down, bundled or discontinued. Holding them "just in case" is a decision with a cost, and we make it deliberately. The accessories category has a specific hazard: device cycles. When a new phone generation launches, cases and screen protectors for the previous one start to age. Stock planning has to know those dates, and range decisions have to accept that some products have a short, predictable life. Fulfilment is part of the product Customers do not separate "the product" from "how it arrived". A well-made cable that takes ten days to arrive in a crushed envelope is, from the customer's point of view, a worse product than an ordinary cable that arrived in two days in sensible packaging. We treat fulfilment as a product surface with its own standards: Dispatch time. A written target from order to hand-off, and a daily check of anything that has breached it. Packaging. Protective enough to prevent damage claims, small enough not to inflate shipping, plain enough not to waste money. Tracking and communication. The customer should never have to ask where their order is. If a shipment stalls, we would rather tell them than wait for the email. Address and delivery exceptions. A clear process for failed deliveries, returns to sender and reshipments, with owners and time limits. Shipping cost is also a pricing decision. Free shipping thresholds, flat rates and carrier choice all change the contribution calculation above. We set them from the numbers, then check whether they changed customer behaviour in the way we expected. Returns: policy, process and learning Returns are unavoidable in consumer products and expensive when handled badly. There are three layers to get right. Policy. The return window, conditions and who pays for return shipping have to be clear, fair and legally compliant in each market we sell to. Consumer protection rules differ between countries, and the policy must meet local requirements; we check those per market rather than copying a template. A generous policy can increase conversion. It can also invite abuse. We decide deliberately and write the reasoning down. Process. A return should move through defined states: requested, approved, received, inspected, refunded or replaced, restocked or written off. Each state has an owner. Refunds go out promptly once conditions are met, because slow refunds generate support tickets, chargebacks and bad reviews. Learning. Every return carries a reason, and reasons are categorised consistently: defect, wrong compatibility, not as described, changed mind, damaged in transit. A cluster of "wrong compatibility" returns usually means product pages are unclear. A cluster of "damaged in transit" means packaging or carrier. A cluster of "defect" on one SKU means a conversation with the supplier. This is exactly the kind of pattern MerchNivo is built to surface for Shopify merchants. Imagine a morning briefing that says: refunds on one product are running well above its usual rate this week, most tagged as compatibility issues, starting after a listing change. The numbers come from the store; the AI only explains them. We describe that approach in An AI e-commerce employee for Shopify stores and the signals themselves in Shopify operations signals. Customer service is a revenue function In a small DTC brand, customer service is often treated as a cost to minimise. We see it differently. The quality and speed of answers affect refunds, chargebacks, reviews and repeat purchase, all of which show up in the numbers. Practically, that means: A shared inbox with clear categories and response-time targets. Written answers for common questions (shipping times, compatibility, warranty, returns) that are kept consistent with the actual policy. A rule that any question asked more than a few times becomes a product page improvement, not just a saved reply. Escalation for anything involving safety, repeated defects or unusual refund requests. Where AI helps draft replies, the facts in the reply (order status, tracking, refund amount, stock) come from the store's systems, never from the model's guess. A human reviews anything that commits the business to money or a policy exception. That is our general principle applied to a support desk. Repeat purchase turns a costly first order into a business Paid acquisition is expensive and getting more so. A DTC brand that relies on every order being a new customer is exposed to every change in ad prices. Repeat purchase is what gives the business room to breathe. Accessories have a natural repeat logic if you look for it. Someone who buys a charger for a new phone may need a case, a second cable for the car, a replacement when the first one frays. Someone who buys a gift may come back for another occasion. The work is to make the next purchase easy and relevant without becoming noise. The levers we use: Post-purchase experience. Good packaging, a clear insert, easy access to the right support content. The first order sets expectations. Relevant follow-up. Messages based on what the customer actually bought, sent only to customers who gave marketing consent. Product-related messages (shipping, warranty) and marketing messages are separate things with separate rules; see Marketing consent vs product consent. Range design. Products that pair naturally, so the store can recommend something genuinely useful. Measurement. Repeat rate and time to second order, tracked by acquisition cohort and channel. We describe the wider thinking in Designing a recurring revenue loop. We avoid tactics that buy a second order at the cost of trust: fake urgency, countdown timers that reset, discount codes so frequent that customers learn never to pay full price. Marketing with the numbers in view Marketing a DTC brand is important. It just comes after the product, the margin and the operations are under control, and it is judged on what it leaves behind. A few operating habits we apply to Noveniq: Spend targets come from contribution, not from revenue. We know how much we can afford to pay for a first order given its likely contribution and repeat behaviour, and we set limits from there. Consistent attribution tagging. Every campaign link follows the group's UTM standards, so channel reports can be compared month to month. Privacy-respecting analytics. We measure what we need to run the business and no more. Creative that matches the product. Claims about compatibility, durability or charging speed must be true and specific. Overpromising in an ad shows up later as returns. When a campaign produces lots of orders with high return rates, it is not a successful campaign. When a channel brings fewer customers who come back more often, it may be the best one we have. The weekly operating review Noveniq has a short weekly review, in the same rhythm as the rest of the portfolio. It covers the same headings each time so trends are easy to spot: Area Questions Sales Orders, average order value, contribution per order versus last period Stock Products below reorder point, stockouts, slow movers Fulfilment Orders past dispatch target, delivery exceptions, damage claims Returns Return rate by product, top reasons, refunds pending Service Open tickets, response times, recurring questions Marketing Spend versus plan, cost per first order, early repeat signals Actions What we are changing this week, and who owns it The review is short because the numbers are prepared automatically and the discussion focuses on exceptions. If everything is inside its normal range, the meeting takes minutes. When something is out of range, that is where the time goes. Common mistakes we try to avoid Most DTC problems are not dramatic. They are small leaks that compound. Ignoring small costs. Payment fees, packaging and a few percent of returns do not look like much on one order. Across a year they can be the difference between profit and loss. Discounting as a habit. A brand that runs a promotion every week trains customers to wait for the next one. Chasing breadth. Adding products because they are available, not because they fit the range and the numbers. Treating stock counts as approximate. Overselling and phantom inventory create refunds, apology emails and lost trust. Letting support answers drift from policy. When the saved reply says one thing and the returns policy says another, customers notice. Scaling spend before fixing operations. More orders through a broken fulfilment process simply produce more broken experiences. Pricing without a race to the bottom Accessories are a crowded category, and the easiest move is to match the cheapest listing you can find. We try not to. Price is set from three inputs: the contribution we need per order, what comparable products of similar quality actually sell for, and what makes the product worth choosing over the cheapest option. That last input is where a brand earns its margin. Clear compatibility information, a sensible warranty, fast dispatch and responsive support are all reasons to pay slightly more, as long as they are real and visible on the product page. When we cannot find a reason for a customer to choose us over a cheaper alternative, that is usually a sign the product does not belong in the range. Price changes are logged with a date and a reason, so that later we can tell whether a shift in conversion or return rate followed a price move, a listing change or something else entirely. How Noveniq fits into the group Noveniq is a separate company inside Oryvelon, with its own customer data and its own accounts. Its customers are not shared with any other product; that rule is described in Shared infrastructure, separate data. It uses group standards for security, domains and analytics, and it goes through the same continue, stop or scale checkpoints as every other company. What flows to the rest of the group is knowledge, not data. When we learn that a certain kind of refund spike usually has a certain kind of cause, that insight can shape how MerchNivo describes signals to other merchants. When we find that a weekly review format works, other companies can borrow it. The customers stay with Noveniq. Summary A DTC brand is an operating business first and a marketing project second. At Noveniq we choose products by landed cost, size, return risk and repeat potential; run the store on contribution margin per order; treat stock as cash and keep one source of inventory truth; hold fulfilment, returns and customer service to written standards; and invest in repeat purchase so the business does not depend on buying every order. Marketing is judged on what it leaves behind, not on revenue. That discipline keeps the store honest, and the lessons it produces make the rest of the group's commerce work better. ## B2B2C licensing to schools: how the model really works URL: https://oryvelon.com/insights/b2b2c-licensing-schools/ Category: Commerce & SaaS Most software has one customer. The person who pays is the person who uses it, or at least works in the same team. B2B2C licensing to schools breaks that. A school or school group signs the contract. A teacher or administrator sets it up. Students use it every day. Parents and guardians want to know what it does with their children's data. Each of those people has different needs, and a product that serves only one of them rarely lasts past the first renewal. EduRelia is Oryvelon's AI learning company for students, families and schools, and school licensing is one of its routes to market. This article explains how we think about the model: who the parties are, how procurement and onboarding work, how data roles are set, and what makes a school renew. Most of it applies to any edtech product selling into schools. Who pays, who decides, who uses The first step is to name the parties clearly, because a sales process that treats "the school" as one person will stall. Party Role What they care about School leadership / trust Budget holder, signs the contract Learning outcomes, cost, risk, reputation IT and data protection lead Approves the product technically Security, data processing terms, integrations, account management Teachers Introduce and supervise use Time saved, fit with curriculum, classroom control Students Daily users Whether it helps, whether it is pleasant to use Parents and guardians Stakeholders, sometimes users Safety, privacy, what their child is actually doing A product can win leadership with outcome claims, lose IT on a missing data processing agreement, and then fail in the classroom because teachers found it added work. Each group has a veto, whether formal or informal. We design the sales materials, the onboarding and the product itself with all five in mind. Why schools buy differently Schools are not slow for no reason. They hold data about children, spend public or tuition money, and answer to parents, inspectors and regulators. Procurement reflects that. Common features of school buying: Budget cycles. Many schools plan spending around an academic or financial year. A product that is perfect in October may have to wait until the next budget. Documentation before demos. Security questionnaires, data protection impact assessments and accessibility statements often come before anyone sees the product in depth. Pilots. Schools frequently want a trial with one year group or one subject before committing. Group decisions. In school groups and trusts, a central team may approve a product once for many schools, or leave it to each school. Low tolerance for surprises. A product that changes behaviour mid-year, or introduces a new data use without notice, can lose trust quickly. The practical response is to prepare the documentation before it is asked for. EduRelia keeps a standing pack: a plain description of what the product does and does not do, a list of the data it processes and why, security controls, data retention periods, subprocessor categories, accessibility information and pricing. When a school asks, we can answer in days rather than weeks. That preparation starts in the company's source-of-truth document, which defines the data rules before any feature is built. Settling data roles before anything else In a school licence, it has to be clear who is responsible for what data, and on what basis. Getting this wrong is a legal problem and, more immediately, a trust problem. In many arrangements under data protection laws such as GDPR and KVKK, the school decides why student data is processed and the provider processes it on the school's instructions. In GDPR terms that usually makes the school the controller and the provider the processor for that data. Other laws, such as COPPA in the United States, have their own rules for children's data and school consent. The exact roles depend on the arrangement and the jurisdiction, and schools and providers should take proper advice; this article describes how we approach it, not what the law requires in any specific case. What that means for EduRelia in practice: A written data processing agreement that states what data is processed, for what purpose, for how long, and what happens at the end of the contract. Processing only on the school's instructions. Student data from a school licence is used to provide the service to that school. It is not used to build a consumer marketing list, train a general model or feed another Oryvelon product. No cross-product use. The group's rule that user data never crosses between companies applies with extra force here. See Shared infrastructure, separate data. Minimal data by default. We collect what the learning experience needs and nothing else. Often that means a pseudonymous student identifier, a class, and learning activity, not full names, dates of birth or home addresses. We explain the reasoning in Child data minimisation in edtech. This is also why a school-licensed student account and a family's direct relationship with EduRelia are kept separate. If a family later chooses to use EduRelia at home under their own account, that is a new relationship with its own consent. It does not inherit the school's data, and the school's data does not become marketing material. The distinction between the two kinds of consent is covered in Marketing consent vs product consent. Tenancy: each school is its own space A school licence implies a boundary. One school's students, teachers and data must never be visible to another school, even within the same trust unless the trust has explicitly asked for shared visibility and is entitled to it. EduRelia treats each licensed school (or trust, where appropriate) as a tenant. Every record carries a tenant identifier, every query is scoped by it, and access checks happen on the server, not just in the interface. Teachers see their own classes. School administrators see their school. Nobody at a school sees anything from another school. We describe the pattern in more depth in Tenant isolation explained. Internally, access follows the same logic. Staff who support schools get the minimum access they need, time-limited where possible, with two-factor authentication and logging. Support staff do not browse student data to answer a question about billing. Pricing a school licence School pricing has to fit how schools budget. Complicated usage-based pricing that produces a different invoice every month is hard for a school to approve, however fair it might be in theory. The structures we consider: Per-student annual licence. Simple to understand and scales with school size. The difficulty is defining "student": enrolled, active, or invited. Tiered school licence. Price bands by school size. Easier to budget, less precise. Per-class or per-subject licence. Suits pilots and gradual rollouts. Group or trust licence. One agreement for many schools, often with a volume price and central administration. Whatever the structure, we try to follow a few principles. The price should be known before the year starts. There should be a clear way to add students mid-year without renegotiation. AI costs should be managed by us, within reasonable fair-use limits, not passed through as unpredictable charges. The wider trade-offs between flat pricing and usage components are discussed in SaaS subscription plus usage. A free pilot can be useful, but it should have a defined scope, a defined end date and agreed success criteria. An open-ended free trial in a school tends to become a permanent free product that nobody is accountable for. Onboarding: where most licences are won or lost A signed contract is not adoption. Plenty of school software is bought, rolled out to a few enthusiastic teachers and then forgotten. The first weeks decide which way it goes. We think of onboarding in three layers. School setup Accounts, classes and rosters need to be created with as little manual work as possible. That might mean a simple upload of class lists, integration with the school's existing identity system, or class codes that teachers share. Whatever the method, the data collected at this stage should match the minimisation rules above: if the product does not need a date of birth, the upload template does not ask for one. Teacher onboarding Teachers need to see, quickly, how the product fits into lessons they already teach. A short guided setup, a handful of ready-made activities mapped to the curriculum, and clear controls over what students can see and do. Teachers also need to understand what the AI does and does not do, and how to report something that looks wrong. That transparency is part of building trust; see Onboarding and trust in AI products. Student and family introduction Students need a first session that works and makes sense without a manual. Families need a plain explanation, usually provided through the school, of what the product is, what data it uses and who to contact with questions. We provide schools with a template letter they can adapt, so the message families receive is accurate and consistent. AI in the classroom: boundaries that schools can check Schools are right to ask detailed questions about AI. Our answers are shaped by the group principle that AI explains, verified data decides. For EduRelia that means: Curriculum content and assessment criteria come from verified sources, not from what a model happens to generate. The AI explains, guides and gives practice. It does not assign official grades or make decisions about a student's placement. Safety hooks run on inputs and outputs, checking for age-inappropriate content and signs that a student may need a human rather than a chatbot. Concerns route to people. If a student writes something that suggests they need help, the product does not try to handle it alone. The school's own safeguarding process is the path, and the product makes that route clear. The general pattern is in Human escalation in AI products. Prompts and student data stay inside EduRelia. The shared AI gateway provides routing, budgets and logging controls, but prompts, knowledge and user data are never merged with other products. We would rather document these limits clearly and have a school ask sharp questions than have them discover something unexpected later. Support that respects the school's time In B2B2C, support requests come from several directions: a student who cannot log in, a teacher who cannot find a class, an IT lead with a security question, a parent with a privacy question. Each needs a clear path. We route most student and teacher questions through the school, because the school knows its own accounts and has the relationship with families. EduRelia supports the school's administrators directly, with defined response times. Privacy and data requests are handled through a documented process that respects the school's role in deciding on them. The metric we watch most closely is support burden per school. A product that generates a steady stream of tickets for the school's IT team is a product that will be questioned at renewal, however much students like it. Renewal is earned during the year Renewal conversations often happen in a few weeks before a contract ends. By then, the decision is mostly made. The school already knows whether teachers used the product, whether students engaged, whether IT had problems and whether any parent complained. What we do throughout the year: Usage reporting to the school. Aggregate, privacy-respecting reports showing how the product is being used: active classes, activities completed, trends over the term. No individual student profiling beyond what the school's teachers need for teaching. Mid-year check-ins. A short review with the school's lead contact: what is working, what is not, what would help. Change notices. Any change that affects data processing, features teachers rely on or pricing is communicated in advance, with time for questions. Clean exit options. Schools should know exactly how to get their data out and how deletion works at contract end. Paradoxically, making exit easy makes renewal easier, because it removes a source of anxiety. See Data retention and deletion. The recurring revenue in a school licence is real, but it is annual and it is earned. What happens at the end of a contract Every school licence ends eventually, whether through non-renewal, a change of provider or a school closing. The end of a contract is also a data event. Our standard approach is written into the agreement: The school is told, in advance, the date on which access will end. The school can export its data in a usable format before that date. After the agreed period, student data from that licence is deleted from active systems, and from backups on their normal rotation cycle. The school receives confirmation that deletion has been carried out. Nothing from a school's data survives in another form: no "anonymised" copy kept for product development unless the agreement explicitly provided for it and it is genuinely non-identifying. A worked example: from pilot to licence Imagine a secondary school that wants to try EduRelia with one year group in one subject for a single term. Here is how we would expect it to run. Before the pilot starts, the school's data protection lead receives the documentation pack and signs a data processing agreement that covers the pilot as well as any later licence. The scope is written down: which classes, which teachers, which start and end dates, and three or four success criteria the school chose itself, such as the share of students completing weekly practice or teacher time saved on preparing revision material. During the term, teachers get a short onboarding session and a named contact. Halfway through, we review usage with the school's lead and fix anything that is getting in the way. At the end, the school receives an aggregate report against its own criteria. Then the school decides. If it continues, the pilot converts into a licence at the price agreed in advance, and the existing classes carry on without a reset. If it does not, the data is exported on request and deleted on the agreed schedule. Either outcome is a clean one. Common mistakes in B2B2C school licensing A few patterns we see often, and try to avoid: Selling to leadership and ignoring IT. The deal stalls at the data protection review. Treating students as the customer. Features built for engagement at any cost can conflict with what the school and families want. Collecting data "for later". Fields added to rosters or profiles without a current purpose become liabilities. Blurring school and home relationships. Using school-provided data to reach families commercially destroys trust fast. Unpredictable pricing. Schools cannot easily approve invoices that vary month to month. Neglecting teachers. If the product adds work to a teacher's week, adoption falls regardless of how good it is for students. Measuring whether a licence is healthy We keep the metrics simple and tied to the parties above: Party Health signal Leadership Renewal intent, satisfaction at check-ins IT / data protection Open questions, incidents (target: none), time to answer requests Teachers Share of licensed classes actively using the product Students Regular use across the term, not just in the first weeks Families Questions and concerns raised through the school These feed EduRelia's own continue, stop or scale checkpoints. If schools sign but classes do not use the product, that is a signal to fix onboarding before signing more schools. Summary B2B2C licensing to schools works when a product serves all its parties at once: leadership who pay, IT who approve, teachers who introduce it, students who use it and families who trust it. For EduRelia that means preparing procurement documentation early, setting data roles clearly in writing, minimising student data, isolating each school as its own tenant, pricing in ways schools can budget, investing heavily in onboarding, keeping AI within boundaries schools can check, and earning renewal through the year with honest reporting and easy exit. None of it is glamorous. All of it decides whether a school stays. ## Unit economics per product: AI costs and shared cost allocation URL: https://oryvelon.com/insights/unit-economics-per-product/ Category: Commerce & SaaS A single product company can get away with rough numbers for a while. Revenue minus costs is profit, and if the bank balance goes up, things are probably fine. A group of companies sharing infrastructure cannot. Once hosting, monitoring, an AI gateway and a handful of tools serve several products at once, it becomes surprisingly easy to lose sight of what any one of them actually costs to run. This article explains how we approach unit economics per product at Oryvelon: how we define a unit for very different businesses, where AI costs belong, and how we allocate shared costs so the numbers help decisions instead of distorting them. It is written for founders and operators, but it applies to any team running more than one product on common foundations. Why per-product unit economics matter in a group Oryvelon builds and operates independent companies, and each one is expected to stand on its own. We explain the principle in Designing companies to stand alone. A company can only stand alone financially if we know, with reasonable honesty, what it earns and what it costs. Three problems appear when that is not measured properly: Cross-subsidy by accident. A product with heavy AI usage quietly runs on costs that the group pays, and looks healthier than it is. Unfair burden. A small product gets charged a share of shared costs that has nothing to do with what it uses, and looks worse than it is. Bad capital decisions. Money and time go to the product with the best-looking spreadsheet rather than the best actual economics. Unit economics are the lowest level at which we can see whether a product works. Everything above that, including monthly company results and portfolio decisions, is built from them. Step one: define the unit A unit is the thing a customer pays for, or the thing that most directly drives revenue. It sounds obvious. In practice, teams often skip it and end up with averages that describe nothing in particular. For the companies in the group, the natural units look like this: Company Natural unit Revenue per unit Main variable costs Noveniq Order Order value after discounts Landed cost, shipping, payment fees, returns MerchNivo Store subscription month Subscription fee plus any usage AI calls for briefings, data sync, support KeşifAtlası Paid report Report price AI explanation, rules lookups, human review time ZodiVela Reading (or credit) Credit price or subscription share AI generation, payment fees EduRelia Licensed student seat per year Licence fee per seat AI usage per student, support per school CastLyra Completed booking or subscription month Commission or fee Verification, moderation, payment handling Some products have more than one unit. KeşifAtlası has free eligibility tests that cost money to run and paid reports that bring revenue in. The free test is not a unit with revenue; it is a cost of acquiring the paid unit, and we treat it that way. The economics of that funnel are discussed in From free test to paid report. The rule we apply: pick the unit that the business will actually be managed on, write it in the company's source of truth, and do not change it lightly. Step two: separate variable, direct fixed and shared costs Once there is a unit, costs fall into three buckets. Variable costs rise with each unit. AI calls per report, shipping per order, payment fees per transaction, verification checks per profile. Direct fixed costs belong to one company but do not rise with each unit. A domain, a product-specific tool subscription, a contractor working only on that company. Shared costs support several companies at once. Group hosting, the AI gateway itself, monitoring, security tooling, shared design and engineering time. Keeping these separate is what makes the numbers readable. Contribution per unit is revenue per unit minus variable costs. It answers the question "does each sale help or hurt?" Company profit subtracts direct fixed costs and an allocated share of shared costs. It answers "does this company pay for itself?" Both questions matter. They are not the same question, and mixing them hides problems. Where AI costs belong AI costs are the part of the model that has changed most in recent years. For some of our products they are the largest variable cost per unit. They are also easy to lose, because model calls are small, frequent and often billed to one account. Our rule is that AI costs belong to the product that caused them, and ideally to the feature and unit as well. The shared AI gateway, described in Building an AI gateway, makes this practical. Every call through it carries a product identifier and a feature tag. That lets us answer questions such as: What does one MerchNivo daily briefing cost in model usage, on average and at the expensive end? How much of a KeşifAtlası report's cost is the explanation, and how much is the follow-up questions? What share of ZodiVela's AI spend goes on readings that are never completed? A shared gateway could, in theory, become a reason to pool AI costs and split them evenly. We think that would be a mistake. The gateway is shared infrastructure; the usage flowing through it is not. Each product has its own budget, its own limits and its own line in the numbers. Costing AI per unit in practice A simple way to cost AI per unit: Tag every model call with product, feature and, where possible, the unit it belongs to (an order, a report, a session). Sum model charges by tag over a period. Divide by the number of units delivered in that period. Track the average and the upper tail, because a small share of very expensive units can distort the average. It is also worth measuring cost per useful outcome rather than per call. A report that needed three retries because output failed schema validation cost three times as much as it should have. See Structured outputs and schema validation for why validation affects cost as well as quality. Levers that change AI cost per unit AI unit cost is not fixed. The main levers: Model routing. Simple tasks go to smaller, cheaper models; hard tasks go to stronger ones. See Model routing. Caching. Repeated explanations of the same rule or the same product data can often be reused safely. See Latency and caching for AI. Prompt size. Sending only the data a task needs reduces cost and often improves quality. Fewer retries. Better prompts and validation reduce the number of calls that fail and repeat. Budgets and limits. Per-product soft and hard limits catch runaway usage before it becomes a large invoice. The wider approach is in Cost discipline for AI products. Allocating shared costs without distorting decisions Shared costs are real, and every company should carry a fair share of them. The difficulty is that any allocation rule is partly arbitrary, and a bad one can make a good product look bad or a weak product look strong. We follow four principles. One written rule, applied consistently. The allocation method is written down and applied the same way each month. It is not adjusted to help a product through a review. Use a driver where one exists. Where there is a clear usage measure, such as storage, requests or compute time, costs are allocated by that measure. Hosting that serves four products can be split by each product's share of requests or resources. Use a simple split where no driver exists. Some shared costs, such as security tooling or group-level design time, have no meaningful usage measure. For these we use a simple, agreed split, which might be equal shares among active companies or a split weighted by stage. Simplicity beats false precision. Show allocated costs separately. In every company report, shared costs appear on their own lines, below contribution and direct costs. Anyone reading the numbers can see what the company controls and what it has been assigned. A worked example Imagine a month in which group hosting costs a certain amount, the AI gateway's own running cost (excluding model usage, which is already attributed) is smaller, and shared tools make up the rest. Four companies are active. Hosting is allocated by each company's share of requests. A product with heavy traffic carries more. Gateway running cost is allocated by each company's share of calls through it. Shared tools are split equally among the four, because they are used roughly evenly and there is no better measure. A new company in its first weeks might carry a small share of hosting (little traffic), a small share of gateway cost (few calls) and a full equal share of tools. That last line can look heavy for a company with little revenue. We accept that, but we read it for what it is: the cost of existing inside the group, not a sign that the product is failing. When to exclude shared costs from a decision For some decisions, allocated shared costs are irrelevant. If we are deciding whether to run a marketing test for Noveniq, the question is whether the extra orders will produce positive contribution. The hosting bill does not change either way. For others they are central. If we are deciding whether to keep a company at all, the question is whether the group would be better off without it, and that includes which shared costs would actually disappear if it closed. Often the honest answer is "not many", which is useful to know. Shared costs that would remain regardless should not be the reason a company is stopped. Unit economics by stage Expectations change as companies move through their life. We would not judge a company in its first month on the same standard as one in its second year. Stage What we expect What we watch Early validation Contribution may be negative while learning Direction of unit cost, willingness to pay First paying customers Positive contribution per unit AI cost per unit, support load per customer Growth Contribution covers direct fixed costs Acquisition cost versus contribution over time Mature Covers allocated shared costs and returns capital Stability of margins, sensitivity to cost changes Negative unit economics early on are acceptable only if there is a credible path to positive ones: a plan to cut AI cost through routing or caching, a price increase supported by evidence, or volume discounts that will actually arrive. "It will get better at scale" is not a plan by itself. These expectations feed directly into the group's continue, stop or scale checkpoints at 30, 60 and 90 days. Acquisition cost and time to pay back Contribution per unit tells us whether each sale helps. It does not tell us whether the business can afford to find customers. For that we compare contribution with acquisition cost. For subscription products such as MerchNivo, the useful question is how many months of contribution it takes to pay back the cost of acquiring a customer, and how many customers stay that long. For order-based businesses such as Noveniq, it is whether the first order plus likely repeat orders covers acquisition. The shape of these calculations differs by model; see SaaS subscription plus usage and Running a DTC brand as a real operating business. Two warnings from experience across the industry. First, acquisition costs rarely stay low as spend increases; the cheapest customers usually come first. Second, repeat and retention assumptions are the easiest numbers to be optimistic about. We prefer to use observed cohort behaviour, even if the sample is small, over hopeful projections. Using unit economics to set prices and limits Unit economics are most useful when they change what we do. Two of the most direct uses are pricing and usage limits. On pricing, the unit view tells us the floor. If a KeşifAtlası report costs a certain amount to produce, including AI explanation, rules lookups and an average share of human review, the price has to sit comfortably above that, with room for payment fees and refunds. It does not tell us the ceiling; that comes from what customers value and what alternatives cost. But knowing the floor prevents a very common mistake: launching at a price that loses money on every sale and hoping volume fixes it. On limits, the unit view tells us where generosity becomes a risk. Imagine a ZodiVela subscription that includes unlimited follow-up questions. If a small group of subscribers asks hundreds each month, their AI cost can exceed what they pay. The answer is not to hide the limit in fine print. It is to design the plan honestly: a clear allowance, a fair-use policy written in plain language, or credits for heavier use. We would rather customers understand what they are buying than discover a surprise restriction. Both uses depend on the same thing: a cost per unit that is recent, measured and trusted by the people making the decision. What unit economics will not tell you Numbers per unit are powerful and incomplete. They say little about brand strength, the quality of customer relationships, regulatory risk or the strategic value of learning in a new market. A product can have modest unit economics today and be worth continuing because it is building something the numbers do not yet capture. We try to be explicit when we make that argument, and to put a date on it. "We accept thin contribution for another quarter because we are testing whether schools renew" is a reasonable position. "The numbers do not matter here" is not. Common mistakes Averaging across products. A group-level "AI cost per user" hides the product that is expensive and the one that is cheap. Ignoring the tail. A few very heavy users or very long sessions can make a subscription unprofitable while the average looks fine. Leaving out human time. KeşifAtlası's escalations to people, CastLyra's manual verification, EduRelia's school support: these are variable costs and belong in the unit. Counting free usage as marketing only. Free tiers and free tests have unit costs. They should be measured and budgeted. Changing allocation rules to suit the story. Once a rule bends for one review, nobody trusts the numbers. Precision theatre. Allocating a small tool subscription by a complex formula wastes time that should go into decisions. The monthly unit economics review Each company produces a short monthly view, prepared from its own data and the gateway's usage records, as part of the group's regular operating rhythm. It covers: Units delivered and revenue per unit. Variable cost per unit, broken into its main parts, with AI shown separately. Contribution per unit and in total. Direct fixed costs. Allocated shared costs, on separate lines. Changes since last month, with a sentence explaining each notable one. One or two actions: a routing change, a price test, a cost to remove. The review is short. Its value is in the trend. A cost per unit that creeps up a little every month is easy to miss in a single snapshot and obvious in a sequence. Summary Unit economics per product start with a clear unit for each company, whether an order, a report, a reading or a school seat. Variable costs, including AI, are attributed to the product and feature that caused them, which a tagged AI gateway makes practical. Direct fixed costs sit with their company. Shared costs are allocated by a simple, written rule, shown separately, and never adjusted to suit a narrative. The result is a set of numbers that tells us honestly whether each sale helps, whether each company pays for itself and where to act next, which is exactly what a group of independent companies needs to make good decisions. ## Verification in talent marketplaces: identity, portfolio and consent URL: https://oryvelon.com/insights/verification-in-talent-marketplaces/ Category: Marketplaces & Consumer Products In most marketplaces, a bad transaction costs money. In a talent marketplace, it can cost much more. A model invited to a fake casting, a creator whose images are reused without permission, a photographer who delivers work and is never paid, a brand that hires someone presenting another person's portfolio. These are not edge cases. They are the reasons many creative professionals are wary of online platforms in the first place. Verification in talent marketplaces is how a platform earns the right to connect strangers. CastLyra is Oryvelon's global marketplace for verified creative talent (models, creators and creative professionals) and the brands that work with them. Verification is not a feature bolted onto it. It is the product's core promise. This article explains how we approach it, and what we think any talent or services marketplace should consider. Why talent marketplaces need more than a badge Many platforms offer a single "verified" badge. It is easy to build and easy to understand. It is also easy to misread. Verified what, exactly? That an email address works? That a phone number received a code? That someone looked at a passport? That the portfolio belongs to the person? A single badge compresses different claims into one symbol and invites users to assume the strongest one. That is a trust problem waiting to happen. If a brand believes "verified" means "this person's portfolio has been checked" when it only means "this person confirmed an email", the badge has made the platform less safe, not more. Our approach is to break verification into distinct checks, run each one properly, and show each one honestly. The checks, separated For talent on CastLyra, the checks fall into four groups. Check Question it answers Typical evidence Contact verification Can we reach this person reliably? Confirmed email and phone Identity verification Is this a real person, and who they say they are? Government ID compared against a live capture, reviewed Portfolio ownership Is the work shown genuinely theirs or their own likeness? Consistency checks, source files, credits, references Consent and rights Do they have permission to show this work here? Declarations, photographer or client permissions where needed For businesses, the checks are different but just as important: Check Question it answers Typical evidence Account verification Is there a real, reachable person behind the account? Confirmed work email and phone Business legitimacy Is this a genuine business or agency? Registration details, website and domain match, public presence Representative authority Is this person entitled to act for the business? Work email on the business domain, confirmation from the business History Has this business behaved well on the platform? Completed bookings, reports, payment record Each check has its own status, its own date and its own rules for when it expires or must be repeated. A profile can be contact-verified and identity-verified while portfolio review is still pending. That is fine, as long as the profile says so. Identity verification without collecting too much Identity verification is the check that people worry about most, and rightly so. It involves sensitive documents. Done carelessly, it creates a store of identity data that is valuable to attackers and unnecessary for the business. The principles we apply: Collect for a purpose, keep for a limited time. The platform needs to know that a person is real and matches their profile. It does not need to hold a copy of their passport indefinitely once that is established. We keep the verification result and the minimum record needed to show that it happened, and delete source documents according to a written retention schedule. See Data retention and deletion. Separate the evidence from the profile. Identity documents are held apart from profile data, with much tighter access. Staff who moderate portfolios cannot open identity files. See Least-privilege access for small teams. Show the result, not the evidence. Brands see that identity has been verified. They never see the document, the document number or the date of birth. Use specialist processes where appropriate. Identity checks involve fraud patterns that are hard to detect by eye. Where a specialist verification process is used, it is chosen for its data handling as much as its accuracy, and it sits behind CastLyra's own boundary. Age deserves particular care. Some creative work involves people under 18, and many legal systems set specific rules for young performers and for their images. Where under-18 talent is involved, a parent or guardian must be part of the account and consent process, and access and contact rules become stricter. The principles behind that are covered in Child data minimisation in edtech, and although that article is about education, the same instinct applies: collect less, restrict more, involve the responsible adult. Portfolio ownership: the hardest check Proving that someone is a real person is comparatively straightforward. Proving that the work they show is theirs is harder. For models and on-camera talent, the question is whether the person in the images is the account holder. For photographers, stylists and other creative professionals, the question is whether they actually made or contributed to the work. For creators, it may be whether the channel or content they point to is genuinely under their control. There is no single test. We combine several signals: Likeness consistency between identity verification and portfolio images, reviewed by a person. Source evidence, such as original files, credits, published work with the person named, or links to channels where control can be demonstrated (for example by posting a short code). References from previous clients or collaborators, where the talent chooses to provide them. Inconsistency flags, such as images that appear elsewhere under another name, or a portfolio whose style and quality shift abruptly. AI can help here, but in a limited way. It can help structure a profile from what the talent provides, suggest missing information and flag possible inconsistencies for a reviewer. It does not decide that a portfolio is genuine or fake. That follows the group principle that AI explains, verified data decides, and in this case the verified data is evidence a person has looked at. Generated profile text is also validated so it never invents credentials, clients or awards the talent did not supply; see Structured outputs and schema validation. Consent is part of verification Creative portfolios contain images of people. Sometimes those people are the talent themselves. Sometimes they are other models in a shoot, or members of the public in the background, or clients whose products appear. Showing those images on a marketplace is a use that needs a basis. CastLyra treats consent as a verification topic, not just a terms-of-service clause: Talent declares rights for each portfolio item: that they appear in it with consent, or have permission from the people and rights holders involved. Photographers and creatives declare that they have the right to display the work, which may depend on their agreements with clients. Takedown is fast. Anyone who appears in an image and did not consent to its use on CastLyra has a clear route to request removal, and requests are handled promptly by a person. Use on the platform is limited to the platform. A brand browsing talent is not being granted a licence to reuse portfolio images. Any use beyond viewing is a matter for a direct agreement. There is also the consent the talent gives to CastLyra itself: to be listed, to be contacted by verified businesses, to receive product messages. That is separate from any marketing consent, and treated separately; see Marketing consent vs product consent. Contact details: released by rules, not by default Verification only protects people if it controls what happens next. A marketplace that verifies everyone carefully and then displays phone numbers and email addresses publicly has done half the job. On CastLyra, talent contact details are not shown on public profiles. Access follows rules: A business must be verified to a defined level before it can send a message or booking request through the platform. Talent choose whether to accept a conversation. Direct contact details are shared only when the talent chooses to share them, or at a defined point in a confirmed booking. Unusual patterns, such as one account messaging many talent in a short time with similar text, trigger review. This design also affects CastLyra's AI features. Talent contact details are kept out of AI prompts entirely, so no model call can accidentally surface them. Trust signals: showing verification honestly Verification is only useful if people can see and understand it. We present trust signals as specific statements, each tied to a check: "Identity verified" with the month it was completed. "Portfolio reviewed" with the month, or "Portfolio review pending". "Business verified" for companies, with the business name as registered. Booking history, shown as counts rather than stars where the sample is small. We avoid signals that sound stronger than the evidence. A profile that has passed contact verification only does not get a badge that implies more. Ratings are only shown once there are enough of them to mean something, and they are tied to completed bookings, not open-ended reviews anyone could leave. We also explain what each signal means in plain language, one click away. Brands and talent should not have to guess what "verified" covers. When verification fails or changes Checks do not last forever, and some fail. Expiry. Identity verification is repeated after a defined period or when key profile details change, such as a legal name. Failed checks. A person whose identity cannot be verified can still hold an account in a limited state but cannot receive booking requests. The reason is explained, and there is a route to try again or to speak to a person. Reports. Any user can report a profile or a business. Reports go to a person, not an automated verdict, and serious reports (suspected impersonation, unsafe requests) are prioritised. The pattern is described in Human escalation in AI products. Revocation. If evidence emerges that a verification was wrong, the status is removed promptly and anyone who interacted with that account in a booking is informed where appropriate. The cost of verification, and who pays Good verification is not free. Identity checks, human portfolio review, business checks and report handling all cost time and money, and they happen before any revenue from that user. Three decisions shape how this is handled: Staged verification. Not every check needs to happen at sign-up. Contact verification comes first; identity before a profile is shown to brands; portfolio review before a profile is promoted or appears in certain categories. That spreads cost and reduces friction for people who are just exploring. Verification as part of the offer. For businesses, verification is part of what gives them access to talent, so it sits naturally inside paid plans or booking fees. How that fits different revenue models is discussed in Marketplace monetisation models. Unit cost tracking. Verification cost per active profile is one of CastLyra's unit economics lines. See Unit economics per product. Verification also slows early growth. A marketplace that verifies strictly will have fewer profiles at launch than one that accepts everyone. We think that is the right trade, and we discuss how to handle it in Solving the cold-start problem. A scenario: a suspicious casting request Imagine a newly registered business on CastLyra that passes account verification and then, within an hour, sends near-identical messages to dozens of models, offering high fees for a shoot in a private location and asking them to continue the conversation on a personal messaging app. What should happen: The volume and similarity of messages trigger a review flag. The business's legitimacy check has not yet been completed, so messages are held rather than delivered in full. A reviewer looks at the account, the business details and the message content. If the business cannot be verified, the account is restricted and the messages are not delivered. Any models who did receive an early message are warned. The pattern is recorded so similar attempts are caught faster. None of this requires a model to make a judgement about intent. It requires clear rules, sensible thresholds and a person with the authority to act. Different talent, different evidence A global creative marketplace covers very different kinds of work, and the same check does not fit all of them. Treating a fashion model, a video creator and a set designer identically either makes verification too heavy for some or too light for others. Models and on-camera talent. Likeness is the core of the work, so the link between identity verification and portfolio images matters most. Measurements and physical details, where talent choose to show them, are self-declared and labelled as such. Creators. Control of the channels they point to is often the most meaningful check. A short code posted to a channel and then removed is a simple, low-intrusion way to show it. Photographers, stylists, make-up artists and other creative professionals. Credits and published work carry more weight than likeness. A named credit in a published editorial or campaign is stronger evidence than an unlabelled image. Agencies representing talent. Here the question shifts to authority: is the agency entitled to represent the people it lists, and do those people know they are on CastLyra? Each represented person should be able to see and control their own profile. Writing these differences down, per category, keeps reviewers consistent. Two reviewers looking at the same kind of profile should reach the same decision for the same reasons. Measuring whether verification works Verification is easy to describe and harder to evaluate. A system that rejects nobody looks efficient and may be useless. A system that rejects many may be protecting people or may be turning away genuine talent. The measures we watch for CastLyra: Measure What it tells us Time from sign-up to each verification level Whether the process is too slow for genuine users Share of profiles that stall at each stage Where friction or confusion sits Reports per active business and per active profile Whether bad actors are getting through Reports upheld after review Whether reporting is being used fairly Verifications later revoked Whether checks are catching what they should Talent-reported unsafe contact attempts The outcome that matters most, which we want as close to zero as possible We review these regularly and adjust thresholds, evidence requirements and reviewer guidance based on what they show, not on a sense that things are probably fine. Common mistakes in marketplace verification One badge for everything. Users read it as the strongest claim. Verifying one side only. Talent-side checks without business-side checks leave the more vulnerable party exposed. Keeping identity documents forever. Once the check is done, the documents become a liability. Publishing contact details. Verification does nothing if anyone can bypass the platform. Letting AI decide. Automated verdicts on identity or ownership are hard to explain and hard to appeal. Hiding what checks mean. Trust signals that nobody understands do not build trust. Summary Verification in a talent marketplace has to run both ways and has to be specific. At CastLyra we separate identity, portfolio ownership, consent and business legitimacy into distinct checks, show each one honestly with a date, keep identity evidence apart and for a limited time, release contact details only by rule, and keep AI in a supporting role while people make verification decisions. It costs more and grows more slowly than accepting everyone. It is also the only way a marketplace for creative talent can deserve the trust of the people on both sides of it. ## Solving the marketplace cold-start problem URL: https://oryvelon.com/insights/solving-the-cold-start-problem/ Category: Marketplaces & Consumer Products A marketplace with no sellers has nothing for buyers. A marketplace with no buyers has nothing for sellers. Each side waits for the other, and an empty platform stays empty. This is the cold-start problem, and nearly every two-sided marketplace has to solve it before anything else matters. It is tempting to treat cold start as a marketing problem: spend enough on acquisition and both sides will arrive. Sometimes that works for well-funded companies in simple categories. More often it produces a large number of sign-ups who look around, find nothing relevant and never return. The empty-room experience is expensive to buy and hard to recover from. This article describes the approach we take to solving the cold-start problem for CastLyra, our marketplace for verified creative talent and brands, and the general principles behind it. The short version: start with supply, go narrow, and match by hand until you understand what you are automating. Why marketplaces start empty Network effects are the attraction of marketplaces. Every new seller makes the platform more useful for buyers, and every new buyer makes it more useful for sellers. Once they are working, they compound. The problem is that network effects run in both directions. At small scale, each side's experience depends on a thin, unreliable other side. A brand searching for a model in a particular city finds three profiles, none suitable. A photographer who joins sees no requests for weeks. Both conclude the platform does not work, and both are right, for them, at that moment. Two things follow. First, what matters is not the total number of users but whether a particular buyer can find a suitable seller when they need one. Second, a platform can be large overall and still empty in the corner a given user cares about. Solving cold start means making specific corners full. Principle one: go narrow before you go wide The most reliable way to beat cold start is to shrink the market until you can fill it. A marketplace that is thin across fifty categories and thirty countries feels empty everywhere. The same number of profiles concentrated in one category and two cities can feel busy. For a talent marketplace, the dimensions to narrow along are: Category. Models, creators, photographers, stylists, make-up artists, video crews. Each has different buyers, different pricing and different expectations. Geography. Much creative work is physical: a shoot happens somewhere. Even with remote work, time zones and language matter. Use case. E-commerce product shoots, campaign work, social content, events. A brand looking for e-commerce models needs different talent from one planning a campaign. Buyer type. Independent brands, larger brands, agencies. They search, decide and pay differently. Picking the first niche is a judgement, not a formula. We look for a combination where supply exists and is underserved by existing options, demand is recurring rather than one-off, the typical booking is valuable enough to justify the effort of matching, and we have some existing knowledge or relationships to start from. Oryvelon's founders have long experience in e-commerce and fashion brands, which makes product and fashion-related creative work a natural candidate for early focus. That is an input to the decision, not a guarantee. Going narrow can feel like giving up on a global vision. It is the opposite. CastLyra is designed as a global marketplace, and every successful global marketplace we can think of started by being dense somewhere. Expanding from a full niche is far easier than filling a thin, wide one. Principle two: start with supply Most marketplaces are easier to start from supply. Buyers arrive with a need and a deadline. If they find nothing, they leave and rarely come back. Supply is usually more patient, especially if the platform gives it value before the first booking arrives. That means asking: why would a talented photographer or model join an empty marketplace? "Because buyers will come eventually" is a weak answer. Stronger answers include: A better profile than they can build elsewhere. A clean, verified portfolio page they can share with their own clients, independent of whether CastLyra sends them work. Verification that means something. For talent who are tired of fake castings, being on a platform where businesses are checked is valuable in itself. See Verification in talent marketplaces. Control over contact. No public phone number, no unsolicited direct messages from unknown accounts. Help with presentation. AI-assisted profile structuring that turns a scattered set of credits and images into a clear page, without inventing anything. Early status. Founding members of a niche can be recognised, within reason, as long as the recognition is honest. This is sometimes called "come for the tool, stay for the network". The tool has to be genuinely useful on its own. If it only makes sense once the network exists, it will not bring supply in. Quality before quantity There is a trap here. It is easy to fill a marketplace with low-quality supply quickly, by importing listings, relaxing verification or recruiting anyone who will sign up. That makes the numbers look good and the product worse. A brand that finds fifty unsuitable profiles is not better off than one that finds none. We would rather launch a niche with a smaller number of carefully verified, well-presented profiles than a large number of thin ones. Verification slows early growth. It also means that every brand who searches finds something credible, which is what brings them back. Principle three: seed demand deliberately Once there is a credible set of supply in a niche, demand has to be brought in on purpose. Waiting for search traffic is too slow at the start. Ways to seed demand for a talent marketplace: Direct outreach to the right buyers. Brands and agencies in the chosen niche, contacted individually, with a specific offer: "we have verified e-commerce models and photographers available in these cities." Solving a real brief. Rather than asking brands to browse, asking them what they need for their next shoot and then finding it. Existing relationships. Where the group already has trusted relationships in a sector, those can open doors, as long as the relationship is honest about what CastLyra is and nothing is shared across companies without permission. User data never crosses between Oryvelon companies; see Shared infrastructure, separate data. Content and search visibility. Guides, rate expectations, briefs and practical information that buyers in the niche actually look for. This takes longer but compounds. See Search and AI visibility for companies. What we avoid: paying for broad demand before supply is ready. A campaign that brings brands to a platform with no suitable talent in their city damages the brand in a way that is hard to undo. Principle four: match by hand, then automate In the first months, the most useful thing a marketplace team can do is make matches personally. A brand submits a brief. Someone on the team reads it, finds suitable talent, checks availability, makes the introduction and follows up after the job. This does not scale, and that is the point. Manual matching teaches you things no dashboard can: What a good brief looks like, and which fields actually matter. Which talent get chosen, and why. Where matches fail: price, availability, style, communication, trust. How long each step takes, and which steps could be removed. What both sides say when things go well and when they go badly. Every one of those lessons becomes a product decision. The brief form gets the fields that predicted good matches. Search filters reflect how brands actually choose. Profile prompts ask talent for the information brands kept asking about. AI features, where they help, are shaped around the real matching process rather than an imagined one. We treat manual matching as research with revenue attached. It is expensive per match, and we measure that cost honestly, as part of unit economics per product. But it is far cheaper than building an automated system around the wrong assumptions. When to reduce manual work Manual matching should shrink as patterns become clear. Signs that it is time: The team can predict which talent a brand will choose from the brief alone. Most briefs follow a small number of shapes. Brands start finding and contacting talent themselves before the team intervenes. The team is spending more time on logistics than judgement. We do not set a fixed date. We watch for these signals and move one step at a time: first a structured brief, then suggested matches reviewed by a person, then self-serve search with support available. Measure liquidity, not sign-ups Sign-ups are the most misleading number in an early marketplace. They measure interest, not function. A platform can add many profiles a week and still fail every brand who searches. The number that matters is liquidity: the likelihood that a participant who comes looking for a match finds one. For CastLyra, useful ways to express it are: Measure What it tells us Share of briefs that receive suitable responses within a set time Whether demand is being served Share of briefs that turn into bookings Whether matches are real Share of verified talent with at least one enquiry in a period Whether supply is getting value Time from brief to booking Whether the process is practical Repeat booking rate among brands Whether the platform is worth returning to These are measured per niche, not across the whole platform. A niche with strong liquidity is ready for more investment. A niche with weak liquidity needs more supply, better matching or a narrower definition before anything else. These measures feed CastLyra's continue, stop or scale reviews, which ask a specific question: is this niche working well enough to expand, or do we need to fix it first? Expanding from a full niche When one niche is working, expansion follows the same logic. Adjacent niches are easier than distant ones: Same category, new geography. E-commerce models in a new city, using the same brief format and verification process. Same geography, adjacent category. Photographers or stylists for the same brands already booking models. Same buyers, new use case. Brands that booked for product shoots now looking for campaign or social content. Each expansion is its own small cold start, but a warmer one. Existing brands can be told about new talent. Existing talent can refer colleagues. The platform's reputation travels ahead of it. We try to expand one dimension at a time. Moving into a new category and a new country at once doubles the uncertainty and makes it hard to tell what went wrong if it does. What never to fake Cold start creates pressure to make a platform look busier than it is. Some marketplaces have filled early listings with invented profiles, scraped listings the owners never agreed to, or fake activity counters. We do not do any of these, and we think they are a mistake even on purely practical grounds. Fake supply fails at the moment it matters: a buyer contacts a profile and nobody answers. Scraped listings put real people on a platform without their consent, which is exactly the harm a verified talent marketplace exists to prevent. Invented activity numbers are claims that can be checked, and a brand that catches one will not trust anything else the platform says. The honest alternative is to be open about being early. A niche page that says, in effect, "we are starting with a carefully verified group of professionals in these two cities, and our team will help match your brief" is more attractive to the right buyer than a page pretending to be a mature marketplace. Early buyers often value the personal attention more than breadth. Keeping early supply engaged Supply is more patient than demand, but not endlessly so. Talent who complete verification and then hear nothing for two months drift away. Keeping them engaged is part of the job. Practical habits that help: Tell talent what is happening. A short, honest update on which briefs are coming in and what kinds of profiles are being chosen. Give profile feedback. When a brand passes over a profile for a fixable reason, such as missing rates or an outdated portfolio, say so privately and constructively. Match fairly. In manual matching, it is easy to keep sending work to the same few people. Spreading early opportunities, where quality allows, keeps more of the supply side invested. Ask for referrals carefully. The best early talent know other good professionals. A personal invitation from a peer carries more weight than any campaign, as long as every referred profile goes through the same verification. Common mistakes in solving cold start Launching everywhere. Thin supply across many categories and places feels empty to everyone. Buying demand before supply is ready. Paid traffic into an empty niche is expensive disappointment. Lowering quality to hit numbers. Unverified or low-quality profiles make every search worse. Treating sign-ups as success. Liquidity is what matters. Automating too early. Building matching algorithms before understanding matches. Charging too early or in the wrong way. A monetisation model that adds friction before value is proven can stop both sides joining. See Marketplace monetisation models. Ignoring the supply side's patience. Talent who join and receive nothing for months will leave, and tell others. How this fits Oryvelon's way of starting companies The cold-start approach is a marketplace-specific version of how we start every company. We validate before we build heavily (How to validate a digital business idea). We often begin with a focused early site that states the offer clearly and collects interest (Phase-zero sites). We work through a defined protocol with written checkpoints (How we start a new company). For a marketplace, validation means testing both sides. Will talent join and complete verification? Will brands submit real briefs? Will a manual match turn into a booking both sides are happy with? A handful of good answers to those questions in one niche is worth more than a large list of interested sign-ups across many. A scenario: the first thirty days in a niche Imagine CastLyra opening a niche for e-commerce product shoots in two cities. In the first weeks, the team recruits photographers, models and stylists who already do this work, helps them build verified profiles, and explains exactly what the platform will and will not do. At the same time, they contact a small number of independent brands that regularly need product imagery and offer to help with their next shoot. Each brief is matched by hand. After each job, both sides are asked what worked and what did not. Patterns emerge: brands care more about turnaround time than the team expected; talent want clearer information about usage rights before accepting. Both become product changes, one a brief field, the other a standard terms section. At the thirty-day checkpoint, the question is not how many people signed up. It is how many briefs became bookings, how quickly, and whether anyone came back. Summary The cold-start problem is solved in small, dense pieces, not across a whole market at once. Narrow to one niche and one geography. Start with supply and give it a reason to join that does not depend on demand already being present, such as a better profile, meaningful verification and control over contact. Seed demand deliberately in that niche. Match by hand until the patterns are clear enough to automate. Measure liquidity rather than sign-ups, and expand one dimension at a time once a niche is working. It is slower than buying growth. It is also how a marketplace becomes somewhere both sides actually want to be. ## Marketplace monetisation models: commission, subscription, lead fees and featured listings URL: https://oryvelon.com/insights/marketplace-monetisation-models/ Category: Marketplaces & Consumer Products Choosing how a marketplace makes money is often treated as a pricing exercise that happens late, once the product exists. It is more fundamental than that. Marketplace monetisation models change who joins, what they expect, how they behave and whether they keep transactions on the platform. A fee structure that looks fine in a spreadsheet can quietly push the best participants away or teach everyone to take the conversation elsewhere. This article compares the main models, commission, subscription, lead fees, featured listings and paid services, looks at the incentives each creates, and explains how we think about the right combination for CastLyra, Oryvelon's marketplace for verified creative talent and brands. The reasoning applies to most services and talent marketplaces. Monetisation shapes behaviour Every fee is a signal. It tells participants what the platform values and what they should do to get the most from it. Charge per transaction, and participants have a reason to transact elsewhere once they have met. Charge per contact, and sellers become wary of any enquiry that might not convert. Charge for visibility, and the most visible listings may be the ones with the most money rather than the best fit. Charge a subscription, and participants start asking whether they got enough value this month. None of these is wrong. Each needs to be chosen with its side effects in mind, and designed so the platform's incentive matches what is good for both sides. A monetisation model that rewards the platform for things users dislike will eventually lose users. Commission: take a share of each transaction Commission is the classic marketplace model. The platform takes a percentage of each booking or sale, from the buyer, the seller or both. Strengths. The platform earns only when a match succeeds, which aligns it with both sides. There is no cost to joining, which helps with cold start. Revenue grows with transaction value. Weaknesses. Commission only works if the transaction happens on the platform. In services and talent marketplaces, that is the hard part. Once a brand has worked with a photographer and has their details, the next booking can happen directly, with no fee. This is often called disintermediation or leakage. What makes commission sustainable. The platform has to keep adding value after the first match. Common ways to do this: Payment protection. Funds held until work is delivered, so talent are paid reliably and brands are protected against non-delivery. Contracts and terms. Standard agreements covering usage rights, cancellation and deliverables. Scheduling and admin. Bookings, reminders, invoices and records in one place. Dispute handling. A person to help when something goes wrong. Ongoing verification. Knowing the other side is still verified and in good standing. If those services are good, many participants will prefer to keep booking through the platform. If they are weak, no amount of rules against going direct will hold transactions in. Commission rates also need to be honest about who bears them. A commission charged to talent, on top of an already competitive creative market, can push rates up for brands or margins down for talent. A commission charged to brands is often easier to absorb as a line item. Subscription: charge for ongoing access In a subscription model, one or both sides pay a recurring fee for access, features or volume. Strengths. Revenue is predictable, which makes planning easier. There is no incentive for participants to take transactions elsewhere to avoid a fee, because the fee is paid either way. Subscriptions suit participants who use the platform repeatedly. Weaknesses. Subscriptions create a barrier to joining, which hurts early growth. They also create a monthly question: "did I get my money's worth?" If a brand pays and finds nothing suitable that month, the subscription feels like a loss. Who should subscribe. The side that receives reliable, repeated value and has a budget for it. In many talent marketplaces, that is frequent hirers: brands and agencies that book talent regularly and value fast access, search tools and account features. Charging talent a subscription simply to be listed is a common pattern elsewhere, and one we are cautious about, because it asks the more vulnerable side to pay before seeing any work. The broader principles of recurring pricing are discussed in Designing a recurring revenue loop and SaaS subscription plus usage. Lead fees: charge per contact or enquiry Lead-fee models charge when a contact is made: a seller pays to respond to a buyer's request, or a buyer pays to reveal a seller's details. Strengths. Simple to understand. Revenue arrives early, at the point of contact, rather than depending on a completed transaction the platform may not see. Weaknesses. Lead fees can create a poor experience on the side that pays. Sellers paying to respond to requests that were never serious, or that were sent to twenty other people, feel exploited. Buyers paying to reveal contacts may find the person unavailable. Lead fees also tend to push the rest of the relationship off the platform, since the platform has already been paid. When they work. Lead fees work best when leads are high quality and scarce, the paying side can judge value before paying, and the platform takes responsibility for lead quality, for example by verifying buyers and limiting how many sellers can respond to one request. For CastLyra, charging talent per enquiry would conflict with our approach to verification and contact rules. We want talent to receive genuine enquiries from verified businesses without paying for the privilege of being asked. Featured and promoted listings Featured listings let participants pay for extra visibility: a higher position in search, a highlighted profile or a place on a category page. Strengths. Easy to add alongside other models. Gives ambitious participants a way to invest in their own growth. Weaknesses. Paid visibility competes with relevance. If search results are ordered by who paid most, buyers learn that the top results are adverts, not the best fit, and trust in search falls. Paid placement can also create an arms race among sellers that mainly benefits the platform. How to do it without damaging trust: Label paid placements clearly, every time. Limit how many paid placements appear in any set of results. Only allow paid placement for participants who already meet the relevance and verification criteria for that search; payment should never place an unsuitable or unverified profile in front of a buyer. Keep organic ranking independent of payment. We are careful with featured listings on CastLyra for all these reasons. If they are used, they sit inside strict rules and never override verification. Paid services and value-added features Many marketplaces earn a meaningful share of revenue from services around the core transaction rather than from the transaction itself. For a talent marketplace, examples include: Enhanced verification for businesses that want to show a higher level of checks. Casting and briefing tools for brands running larger projects. Portfolio presentation features for talent, kept optional and never required for being contacted. Managed matching, where the platform's team handles a complex brief end to end. Usage rights and contract templates for specific kinds of work. Strengths. Revenue comes from people who value the extra service, so it does not add friction for everyone. Services often make the core marketplace work better. Weaknesses. Each service has its own costs and operations. Too many add-ons make the offer confusing. And any paid feature for the vulnerable side should be genuinely optional, never a hidden requirement for fair treatment. Comparing the models Model Who usually pays Aligned with successful matches? Risk to trust or growth Revenue predictability Commission Buyer, seller or both Yes Leakage off-platform Varies with volume Subscription Frequent users, often buyers Partly Barrier to joining High Lead fees Usually sellers Weakly Poor lead quality, resentment Medium Featured listings Sellers No, unless tightly bounded Relevance and trust erosion Medium Paid services Whoever values the service Often Complexity, hidden requirements Medium Most successful marketplaces combine two or three models. The combination matters more than any single choice. Timing: when to start charging Monetisation interacts directly with cold start. A marketplace that charges both sides from day one may never get enough participants to become useful. A marketplace that stays free too long may teach users that the service has no value, and find it hard to introduce fees later. We discuss the early stage in Solving the cold-start problem. A pattern that often works: Free for the harder-to-attract side from the start, especially supply. Charge the other side early for clear value, such as a completed, well-matched booking, even if the price is modest. Introduce subscription or paid services once repeat users exist and can see what they get. Add visibility products last, if at all, once search quality is well established and can be protected. Whatever the timing, fees should be announced before they apply, with a clear explanation of what they pay for. The CastLyra approach For a verified talent marketplace, our reasoning leads to a particular combination. Brands and businesses carry most of the revenue. They receive the most concentrated value (access to verified talent, a safer hiring process, admin handled in one place) and they usually have a budget line for creative production. In outline: Talent join, verify and receive genuine enquiries without paying. Paying to be contacted would undermine the protection verification offers, and would weigh on the side most exposed to bad actors. Businesses pay for access and completed bookings, through a combination of a booking fee and, for frequent hirers, a subscription that covers search tools, briefing features and account management. The balance between these is something we expect to tune based on evidence. Payment protection and standard terms justify keeping bookings on the platform, rather than rules alone. Optional paid services for both sides, never required for fair treatment. Promoted visibility, if used at all, is tightly limited, clearly labelled and only available to profiles that already meet verification and relevance criteria. This is a design position, not a set of results. Pricing levels and the mix between fees will be set and adjusted using real booking data and the unit economics of verification, matching and support; see Unit economics per product. Changes go through CastLyra's continue, stop or scale reviews rather than being made on instinct. Costs that monetisation has to cover A monetisation model is only right if it covers what the marketplace actually costs to run. For a talent marketplace, those costs are heavier than they look: Verification: identity checks, portfolio review, business checks, re-verification. Moderation and safety: handling reports, reviewing suspicious activity, removing bad actors. Payments: processing fees, payment protection, refunds and disputes. Support: helping both sides before, during and after bookings. Matching: manual matching in early niches, and the tools that replace it later. AI usage: profile structuring and brief assistance through the group's shared AI gateway, attributed to CastLyra's own budget. See Cost discipline for AI products. If the model only covers transaction processing and ignores verification and safety, the marketplace will be tempted to cut exactly the parts that make it trustworthy. Handling leakage without heavy-handed rules Every services marketplace faces the question of what to do when participants take a relationship off the platform. Some respond with strict rules: contact details blocked from messages, penalties for mentioning a phone number, contract clauses forbidding direct work. Some rules are reasonable. On CastLyra, contact details are released by rule for safety reasons, not to trap anyone, and early messages are kept on the platform so that reports and reviews have a record to work from. But using rules as the main defence against leakage tends to fail. Determined participants find ways around them, and honest participants feel policed. A better defence is to make staying worthwhile. Imagine a brand that has booked the same photographer three times. If booking a fourth time through CastLyra means payment protection, a standard contract with usage rights already written, automatic invoices and a record of every job in one place, many brands will keep using it. If it only means a fee, they will not. Two further habits help. First, price the ongoing relationship fairly: a lower fee for repeat bookings between the same parties can keep them on the platform at a small cost. Second, accept some leakage as normal. A marketplace that introduces a brand to a long-term creative partner has created real value, even if not every later booking passes through it. The goal is a model that stays healthy with realistic leakage, not one that depends on eliminating it. A worked example: what one booking has to pay for Consider a hypothetical booking in which a brand hires a photographer for a one-day product shoot. The numbers below are illustrative only, not CastLyra pricing. Line Example Photographer's fee, agreed between the parties 1,000 Booking fee paid by the brand (10%) 100 Payment processing and protection costs −30 Share of verification and re-verification costs −15 Share of moderation, safety and support −20 AI usage for brief assistance and profile structuring −3 Contribution from the booking 32 Even in a simplified example, most of the fee goes on the things that make the booking safe and reliable. If the booking fee were lower, or if a large share of repeat bookings moved off the platform, contribution could disappear quickly. That is why the monetisation model has to be designed together with the cost of trust, and why subscription revenue from frequent hirers can make the whole model steadier. The same kind of calculation, run with real numbers per niche, tells us whether fees are set sensibly, whether verification cost per booking is falling as the platform matures, and whether subscriptions are covering what they should. Monetisation across Oryvelon's other products The same principle, that the revenue model should reward the behaviour you want, shows up across the group in different forms. KeşifAtlası offers a free eligibility test and a paid report, so revenue depends on the report being worth more than the free answer; see From free test to paid report. MerchNivo uses a subscription because the value arrives every day. CastLyra's marketplace model is different again, because two sides with different needs have to be served at once. Common mistakes in marketplace monetisation Charging the fragile side first. Supply leaves before demand arrives. Commission without value. Participants go direct after the first match, and the platform tries to stop them with rules instead of services. Pay-to-rank search. Buyers stop trusting results. Lead fees on unverified leads. Sellers pay for noise and resent it. Too many add-ons too early. The offer becomes confusing before the core works. Changing fees without notice. Trust lost in a pricing change is slow to rebuild. Ignoring the cost of trust. Pricing that does not fund verification and safety erodes them over time. Summary Marketplace monetisation models are design decisions that shape behaviour. Commission aligns the platform with successful matches but needs real ongoing value to prevent leakage. Subscriptions give predictability and suit frequent users who get repeated value. Lead fees work only when leads are high quality and the platform stands behind them. Featured listings need strict labelling and limits so they never override relevance. For CastLyra, that points to a model weighted towards the businesses that hire: talent join and get contacted without paying, businesses pay through booking fees and subscriptions, optional services add value, and any paid visibility stays tightly bounded. The mix will be tuned by evidence, and it has to fund the verification and safety that make the marketplace worth using. ## Rules engines for eligibility: why AI explains and verified rules decide URL: https://oryvelon.com/insights/rules-engines-for-eligibility/ Category: Marketplaces & Consumer Products Ask a general-purpose chatbot whether you qualify for a student visa and you will usually get a confident, well-written answer. It may even be right. The trouble is that you cannot tell which parts are right, which rule it relied on, or whether that rule was still in force last month. For a question with real consequences, fluent is not the same as correct. That gap is why KeşifAtlası, our visa and relocation eligibility company, is built on a rules engine for eligibility. Verified rules decide the outcome. AI explains the outcome. People handle the questions neither should answer alone. This article describes how we think about that design — what a rule is, how it is stored and versioned, what AI is allowed to do around it, and what happens when an official rule changes overnight. The pattern applies well beyond visas: benefits, grants, scholarships, insurance cover, loan pre-checks and programme admissions all have the same shape. What an eligibility rules engine actually is At its simplest, a rules engine takes two inputs — a set of facts about a person and a set of rules — and returns an outcome with reasons. The facts are answers: nationality, age band, purpose of stay, education level, whether the person holds an offer letter, how much they can show in funds. The rules are conditions: "applicant must hold an acceptance letter from a recognised institution", "applicant must be able to show a stated minimum of funds", "applicants of these nationalities must apply from their country of residence". The engine evaluates each relevant condition and returns something like: Likely eligible — all required conditions met on the answers given. Not eligible on current answers — one or more hard conditions fail, with the specific reason. Needs review — the answers touch a condition that is ambiguous, discretionary or outside what the engine can judge. That third outcome matters as much as the first two. A good eligibility system knows the edge of its own competence. We come back to it below. What the engine is not: it is not a model reasoning its way to a conclusion, and it is not a set of if-statements scattered through application code. It is a deliberately separate layer, with its own data, its own review process and its own release history. This is the concrete version of a principle we apply across the group, described in AI explains, verified data decides. Rules are data, not code The most important design choice is to treat rules as structured data with provenance, not as logic embedded in the product. Each rule in a well-designed eligibility database carries roughly these fields: Field Purpose Rule ID A stable identifier that never changes, even when the rule's content does Jurisdiction and route Which country and which visa or permit category the rule belongs to Condition The machine-evaluable test, written against named input fields Plain-language statement The rule as a person would read it, reviewed by a human Source The official page, regulation or guidance the rule comes from Effective from / to The dates the rule applies to Version Incremented every time the rule's meaning changes Confidence type Hard requirement, typical requirement, or discretionary factor Owner and reviewer Who wrote it and who checked it Last verified When someone last confirmed it against the source Two of these deserve a comment. Confidence type separates rules that are binary from rules that are not. "Must hold a passport valid for the length of stay" is a hard requirement. "Officers may consider ties to the home country" is a discretionary factor. The engine should never turn a discretionary factor into a pass or fail. It can flag it, explain it and route it to review. Last verified is different from effective from. A rule can be in force for years; we still want to know that someone checked it recently. A stale verification date is a signal to re-check, not a reason to stop answering — but it should be visible internally. Storing rules this way has practical benefits. Non-engineers can review them. Changes produce a readable diff. The same rule can be shown to a user in plain language and evaluated by the engine without two copies drifting apart. And the rules database becomes an asset the company owns, independent of any model provider. Why a language model should not decide It is tempting to skip the rules layer and let a capable model read official guidance and answer directly. We decided against that for KeşifAtlası, for four reasons. Reproducibility. The same person, asking the same question twice, should get the same answer. A rules engine guarantees that. A model, even at low temperature, does not — and small wording changes in the question can shift the answer. Traceability. When someone asks "why am I not eligible?", the honest answer is a specific rule with a specific source. A model can generate a plausible reason, but that reason is not guaranteed to be the one that actually drove the output. Currency. Official rules change. A model's training data has a cutoff, and retrieval over web pages can surface an outdated page as easily as a current one. A versioned rules database with effective dates knows which rule applies on which date. Accountability. When a rule is wrong in our database, we can find it, fix it, and identify every assessment it affected. When a model's reasoning is wrong, there is nothing concrete to fix. None of this means AI is useless here. It means AI has a different job. What AI does do: explanation, intake and translation Inside KeşifAtlası, AI works around the rules engine rather than inside it. Its jobs are the ones language models are genuinely good at. Explaining outcomes. The engine returns structured results: rule IDs, pass or fail, the input values that mattered. The model turns that into a clear, calm explanation: what the result means, which requirement is not met, and what the person could look into next. It is given the plain-language rule statements and sources as context, and its output is checked against a schema so it cannot introduce a requirement that is not in the result. See Structured outputs and schema validation. Helping with intake. People describe their situation messily. "I finished a two-year programme but it's not really a degree" needs to become a structured answer about education level. AI can suggest the mapping and ask a clarifying question; the person confirms. The engine only ever evaluates confirmed answers. Translation and tone. KeşifAtlası starts from Türkiye, so many users will read results in Turkish while the official sources are in another language. The model helps produce readable explanations, while the rule statements themselves are human-reviewed. Answering follow-up questions — within limits. A person can ask "what counts as proof of funds?" and get an answer grounded in the rule's reviewed text. If the question goes beyond what the rules database covers, the model says so rather than improvising. The boundary is simple to state: the model may explain a decision; it may not make one. If an explanation and the engine's result ever disagree, the engine's result is the one shown, and the mismatch is logged as a defect. Designing the questionnaire: the free test The rules engine is only as good as its inputs. In practice, most of the product design effort goes into the questions. A few principles we apply: Ask only what a rule needs. Every question must map to at least one rule condition. If no rule uses it, we do not ask it. This keeps the test short and keeps sensitive data to a minimum, which matters a great deal for a product that touches nationality, finances and family situation. Use bands where exact values are not needed. If a rule only cares whether funds exceed a threshold, ask the band, not the exact balance. Branch early. Purpose of stay and nationality determine most of the rest. Ask them first and skip everything that cannot apply. Allow "not sure". A person who does not know whether their qualification is recognised should not be forced to guess. "Not sure" routes that condition to review instead of producing a false fail. Show progress honestly. People abandon forms that feel endless. A visible, accurate step count helps. This short questionnaire is the free test at the top of KeşifAtlası's funnel. How that connects to the paid report is covered in From free test to paid report. Versioning: every answer knows its rule set Here is the design detail that separates a trustworthy eligibility product from a quiz: every assessment stores the rule-set version that produced it. When a person completes the test, we record their confirmed answers, the rule-set version, the outcome and the list of rules that fired. Six months later, if they ask why their report said what it said, we can reproduce it exactly — even if the rules have changed three times since. This also means we can answer a harder question: which past assessments would change under the new rules? When a funding threshold goes up, we can identify reports produced under the old threshold where the person's answers sat between the old and new figures. Whether and how to notify those people is a product and consent decision — see Marketing consent vs product consent — but the ability to find them comes from versioning. The same idea underpins how we treat prompts: explanation prompts are versioned too, and a report records which prompt version wrote its explanation. See Prompt versioning and evaluation. When the official rules change Rules change without warning. A ministry updates a funding requirement. A new route opens. A document requirement is added for one nationality. For an eligibility product, this is not an edge case; it is normal operations. We treat a rule change as a release, with the same discipline as a code release. Detect. Someone responsible for a jurisdiction watches its official sources. Monitoring can help flag that a page changed, but a person decides whether the rule changed. Record. The reviewer writes the new version of the rule: new condition, new plain-language text, source link, effective date. The old version is closed with an end date, not deleted. Review. A second person checks the new rule against the source. For a small team this may be the other founder or a trained reviewer; the point is two sets of eyes on anything that changes an outcome. Test. The rule set is run against a library of saved example cases — synthetic profiles covering typical and edge situations, never real users' data. Every case whose outcome changes is inspected. An unexpected change usually means the new rule was written too broadly. See Environments and test data. Release. The new rule-set version goes live with its effective date. The product shows the date rules were last updated for that route. Follow up. Affected past assessments are identified. Support content and explanation context are refreshed. A worked example Imagine a destination country announces that, from the first of next month, the minimum funds required for a particular study route will rise. The jurisdiction reviewer sees the announcement on the official site. They create version 4 of rule STUDY-FUNDS-01, with the new threshold, a link to the announcement, and an effective date of the first of next month. Version 3 gets an end date of the last day of this month. A second reviewer confirms the figure and the date. The test library runs. Twelve synthetic cases change outcome, all of them profiles with funds between the old and new thresholds — exactly as expected. One additional case changes, a profile on a different route. Investigation shows the condition was accidentally attached to both routes. The reviewer fixes it, re-runs, and only the expected twelve change. On release day, anyone taking the test for a start date after the first of next month is assessed against version 4. Anyone whose planned start date is earlier is assessed against version 3 — because effective dates are evaluated against the person's intended timeline, not today's date. Existing reports show a short note that the rule has since changed. No model retraining. No prompt rewrite. No guesswork about which answers were affected. Routing hard questions to people Some questions do not belong to software at all. Discretionary decisions, unusual immigration histories, complex family situations and anything that looks like legal advice should reach a qualified person or be clearly declined. KeşifAtlası's engine has three ways to trigger this: A discretionary rule fires. The outcome becomes "needs review", with an explanation of why. The person answers "not sure" on a condition that matters. Review rather than a guessed result. The follow-up conversation goes out of scope. The explanation layer detects a question it cannot ground in the rules and offers a route to a person instead of an answer. Designing these hand-offs well — what the person sees, how quickly they hear back, what context travels with them — is its own topic. See Human escalation in AI products. We are also explicit in the product that an eligibility result is an informational assessment, not a legal opinion or a guarantee of a visa decision. Official authorities decide applications. Our job is to help people understand where they stand and what to prepare. Common mistakes in eligibility products Having looked at how eligibility tools tend to go wrong, a few patterns come up repeatedly. Hard-coding rules in the front end. It works for the first version and becomes unmaintainable the first time a rule changes. Nobody can see what the product actually believes. Deleting old rules. If the old version is gone, past answers cannot be explained. Close rules with an end date; never overwrite them. Treating "not eligible" as the end. A person told they fail one requirement wants to know which one and whether anything can change. A good result names the rule and, where appropriate, what the person could look into. Mixing sources of different authority. Official regulations, embassy guidance and community forum advice are not equal. Only the first two belong in the rules database, and the source field makes the difference visible. Letting the explanation layer add requirements. A model asked to "be helpful" will sometimes add sensible-sounding extra documents. Schema validation and a check that every requirement mentioned maps to a fired rule stop this. Collecting more data than rules need. Every extra question is extra sensitive data to protect and extra friction. See Data retention and deletion. How this fits the rest of the group KeşifAtlası uses the group's shared AI gateway for its explanation layer: model routing, budgets, safety hooks, schema validation and logging. See Building an AI gateway. What it does not share is its rules database, its users or its prompts. Those belong to KeşifAtlası alone, in line with Shared infrastructure, separate data. The same pattern — verified system decides, AI explains — shows up elsewhere in the portfolio. MerchNivo takes numbers from the Shopify store, never from a model. CastLyra takes verification status from its verification records. Eligibility is simply the case where the stakes are most personal and the rules change most often, which makes the discipline most visible. A short checklist for building your own If you are designing an eligibility product — visas, grants, benefits, admissions — these are the questions we would ask first: Is every outcome produced by a written rule with a source, not by a model? Does every rule have a stable ID, a version and effective dates? Can you reproduce any past assessment exactly? Is there a "needs review" outcome, and does it route somewhere real? Does every question map to at least one rule? Is there a saved test library that runs on every rule change? Does the explanation layer have a schema that prevents it from adding requirements? Does the product say plainly that the result is informational and who actually decides? If the answer to any of these is no, that is where to start. Summary A rules engine for eligibility separates two jobs that are easy to blur: deciding and explaining. At KeşifAtlası, verified, versioned rules — each with a source, effective dates, an owner and a reviewer — decide whether a person looks eligible, not eligible or in need of review. AI turns that structured result into a clear explanation, helps with messy intake and translation, and declines questions it cannot ground. Every assessment records the rule-set version that produced it, so past answers can be reproduced and affected reports can be found when rules change. Rule changes are handled as releases: detected, recorded, reviewed, tested against a library of synthetic cases and released with effective dates. The result is an eligibility product that can be trusted precisely because it can show its work. ## Positioning consumer AI as reflection and entertainment, honestly URL: https://oryvelon.com/insights/entertainment-positioning-consumer-ai/ Category: Marketplaces & Consumer Products Astrology, tarot, numerology and traditional practices like coffee-cup reading have been part of how people reflect on their lives for a very long time. Many people enjoy them without believing they predict anything. They like the ritual, the language, the chance to think about themselves from an unfamiliar angle. That is a real need, and a well-made product can meet it. It is also a category full of products that overpromise. "Your love life will change on Thursday." "The stars say you should quit." Add a generative AI model that can produce endless confident text, and the temptation to overpromise gets stronger. ZodiVela, our personal discovery company, takes the opposite position. It is designed and described for reflection and entertainment, not prediction or advice. This article explains what that positioning means in practice for a consumer AI product — in copy, in prompts, in safety design, in pricing and in the visual experience — and why we think honest positioning makes a better business, not a smaller one. Positioning is a product decision, not a disclaimer The weakest version of entertainment positioning is a line of small print: "For entertainment purposes only." The product then goes on to promise, predict and pressure, and the disclaimer does nothing except protect the company on paper. We treat positioning as something that has to show up in every layer of the product: Layer What honest entertainment positioning changes Brand and copy Describes readings as reflection, ritual and discovery; never promises outcomes Prompts Instructs the model to write in reflective, open language and forbids predictions and instructions Output validation Checks responses for predictive claims, certainty language and advice before they are shown Safety hooks Detects health, legal, financial and crisis topics and changes the product's behaviour Features Excludes features that only make sense if readings were predictive, such as "best day to invest" Pricing and prompts to buy Avoids fear, urgency or implied consequences for not buying Support Handles complaints and refund requests with the same framing If the positioning only exists in the footer, it is not positioning. It is a liability shield. ZodiVela's source-of-truth document states it plainly under "will not do": readings will not be presented as predictions or as medical, legal or financial advice. See Why every company starts with a source-of-truth document. Reflection language versus prediction language The most practical tool we have is a clear distinction between two kinds of sentence. Prediction language tells the reader what will happen or what they should do. - "You will meet someone important this month." - "Avoid signing contracts until the full moon." - "This card means your relationship is ending." Reflection language offers a theme and invites the reader to connect it to their own life. - "This period is traditionally associated with new connections. Where in your life are you open to meeting someone new?" - "Many readers use this phase to review commitments before making them. Is there anything you'd like more time to think over?" - "In tarot tradition, this card is linked to endings and transitions. What might be ready to close, and what might open?" The second set is not vaguer or less interesting. It is often more engaging, because it asks the reader to do something with it. It respects the tradition's symbolism while leaving the decision where it belongs: with the person. This distinction goes into ZodiVela's writing guidelines, its prompts, and its evaluation set. When we test a new prompt version, one of the checks is simply: does any output contain a prediction or an instruction? See Prompt versioning and evaluation. What the AI does and does not decide Our group-wide principle is that AI explains and verified data decides. In an entertainment product, that principle takes a slightly different form, but it still holds. The structured part of a reading comes from deterministic sources. A birth chart is calculated from date, time and place using established astronomical methods. A numerology number is computed from a name or date by fixed rules. A tarot spread is drawn by a random process and recorded. The symbolism attached to each placement, number or card comes from a curated, human-reviewed library rather than being invented fresh each time. The language part — turning those elements into a flowing, personal, readable interpretation — is where the model helps. It is given the calculated elements and the curated meanings, and asked to write a reflective reading within strict rules. This split matters for two reasons. First, consistency: two people with the same chart placement see interpretations drawn from the same meaning, not from whatever the model felt like generating. Second, honesty: the product can say exactly what is calculated and what is interpreted. We explain the principle more generally in AI explains, verified data decides. Safety hooks for sensitive moments People bring real life to a reading. Someone asks about a health scare, a court case, a pregnancy, debt, a breakup that has left them in a very dark place. A product that answers those questions with a cheerful card interpretation is doing harm, however good its disclaimer. ZodiVela's safety design has three levels. It runs through the group's AI gateway, which provides safety hooks as a shared service while ZodiVela owns its own policies. See Safety hooks for consumer AI products. Topic detection before generation. Questions are checked for health, legal, financial and crisis topics before the reading is written. When one is detected, the product changes what it does. For health, legal or money questions, the reading steps back: it acknowledges the question matters, says clearly that a reading is not the place to decide it, and suggests speaking to a qualified professional. It can still offer a reflective reading on a related, non-advisory theme if the person wants one. Crisis handling. Signals of self-harm or immediate danger stop the reading entirely. The person sees a calm, direct message and information about where to get immediate help. No upsell, no "try another card". Output checking after generation. Even with good prompts, a model may slip into certainty or advice. Outputs are checked for prediction phrases, instructions and sensitive claims. Anything that fails is regenerated or replaced with a safe fallback. See AI fallbacks and degraded modes. These rules are written by people, reviewed, and tested with a library of example questions. They are not left to the model's judgment. Designing for delight without manipulation Being honest does not mean being dull. A reflection product lives or dies on whether people enjoy it, and enjoyment is a design problem. What we think creates delight in this category: Craft. Beautiful typography, considered illustration, calm colour, good pacing. A reading should feel like opening something made with care. Personal relevance. The reading uses the person's own chart, numbers or spread, and connects themes to things they have chosen to share. It should feel written for them, because in a real sense it is. Ritual. Drawing a card, waiting a moment for a reveal, returning to a daily reflection. Small rituals make the experience memorable. Depth on demand. A short daily reflection for someone with a minute; a detailed, sectioned reading for someone who wants to spend twenty. Journalling. A place to note what a reading brought to mind. The value is the person's own reflection, which also makes the product more useful over time. What we deliberately exclude: Fear. "Something is blocking your energy — buy the full reading to find out what." Manufactured urgency. Countdown timers on readings, "only today" framing for ordinary features. Implied consequences for not paying. Suggesting that skipping a paid reading leaves the person exposed to a risk. Endless compulsion loops. Notification patterns designed to create anxiety rather than invite reflection. These tactics can lift short-term conversion. They also erode trust in a category where trust is already fragile, and they attract exactly the refunds, complaints and platform reviews that damage a consumer brand. Pricing an entertainment product honestly ZodiVela earns through credits and subscriptions. Both can be run honestly or manipulatively, and the difference is mostly in the details. Credits suit people who want an occasional detailed reading. The honest version shows the price in credits and in money before purchase, never hides what a reading includes, and does not let credit packs be structured so that people always end up with an unusable remainder. Subscriptions suit people who value a daily or weekly reflection habit. The honest version makes the renewal date and price obvious, makes cancellation as easy as sign-up, and does not bury the subscription behind a free trial that converts silently. The free tier matters too. A free daily reflection or a short first reading lets people see the quality and tone before paying anything. That is covered in more detail, alongside KeşifAtlası's model, in From free test to paid report. The economics of AI generation apply here as well. Detailed readings cost more to generate than short ones, and a subscription with unlimited long readings can lose money on heavy users. Caching curated meanings, routing short reflections to smaller models and setting fair-use limits keep the unit economics sound. See Cost discipline for AI products. Data: personal but not over-collected Readings are personal, and some inputs are sensitive. A birth date, time and place is enough to calculate a chart; a name enables numerology; a photo of a coffee cup is an upload the person has chosen to share. Questions people type can reveal a lot about their lives. Our rules for ZodiVela follow the group's data principles: Collect what a reading needs. A birth time is useful for a chart; it is optional, and the product explains what changes without it. Keep readings and questions inside ZodiVela. No cross-product profiling. A person's reading history is never used by another Oryvelon company. Treat uploads carefully. Cup photos are used to produce the reading and handled under clear retention rules. See Data retention and deletion. Let people delete. Readings, journals and account data can be removed. Worked example: a question the product should not answer Imagine a user opens ZodiVela late at night and types: "Should I put my savings into my friend's business? Pull a card for me." A product built for engagement at any cost would pull a card and write something like "The Wheel of Fortune says luck is on your side — trust your instincts." That reads as financial advice from a card. ZodiVela's flow works differently. The question is checked before generation and flagged as a financial decision. The response acknowledges that this is an important choice, says clearly that a reading cannot and should not decide it, and suggests talking it through with someone qualified to look at the numbers. It then offers, if the person still wants it, a reflective card on a related theme — for example, on trust and boundaries in friendships — written without any reference to whether to invest. The person still gets something of value: a moment of reflection and an honest steer. The company avoids a response it would be embarrassed to see screenshotted. Both outcomes matter. Common mistakes in entertainment AI products Relying on the disclaimer. If copy, prompts and features all imply prediction, a footer line does not change what users experience. Letting the model choose meanings freely. Without a curated meaning library, interpretations drift, contradict each other and occasionally say something harmful. Treating safety as moderation only. Filtering bad words is not enough. The product must behave differently when a question is about health, money, law or crisis. Optimising for session length. Reflection products that chase time-on-app tend to become anxious products. A good daily reflection can take ninety seconds. Hiding the price or the cancellation. Consumer trust in this category is low. Clear pricing is one of the few ways to stand out. Over-collecting data "for personalisation". Personal does not mean more data. It means using the data people chose to share, well. Why honest positioning is also good business It is fair to ask whether all this restraint costs revenue. In the short term, some of it probably does. We still think it is the right commercial choice. Honest positioning reduces refunds and chargebacks, because people get what they were told they would get. It reduces platform and payment-provider risk, since app stores and payment providers look closely at manipulative subscription patterns. It makes the brand shareable in a way that fear-based products are not; people recommend things that made them feel good. And it gives ZodiVela a durable place in a crowded category: the reading app that is beautiful, personal and straight with you. It also keeps ZodiVela consistent with what Oryvelon is. Each company stands alone with its own brand, but the "An Oryvelon company" endorsement only works if every company behaves in a way the group is willing to stand behind. See Naming, domains and brand architecture. How we test the positioning before launch Positioning is easy to agree with in a meeting and easy to lose in production. Before a new reading type or prompt version goes live, we test it against the promise. An evaluation set of real-shaped questions. A library of example questions, written by the team, covers ordinary curiosity ("what does my rising sign mean?"), borderline topics ("will my exam go well?") and clearly sensitive ones (health, money, law, crisis). Every prompt version is run against it. Three checks on every output. Does it contain a prediction? Does it contain an instruction the person is meant to follow? Does it respond correctly to a sensitive topic? A single failure on a crisis example blocks release; failures elsewhere are reviewed and fixed. A read-aloud test. Someone on the team reads a sample of outputs out loud. Language that sounds fine on screen — "you must", "this is certain", "the cards warn you" — is much easier to catch when spoken. A copy review against the source of truth. Store listings, ads, onboarding screens and email subject lines are checked against the "will not do" list. Marketing copy is where overpromising most often creeps back in, because it is written fastest. The onboarding moment matters especially, since it sets expectations for everything after it; see Onboarding and trust in AI products. A positioning checklist for reflection products If you are building in this category, these are the questions we would ask of any screen, prompt or campaign: Would a reasonable reader take this as a prediction of a specific event? Does it tell the person what to do about health, money, law or relationships? Is the calculated or drawn part of the reading separated from the interpreted part? Does a sensitive question change what the product does, not just what it says? Is the price shown in money before purchase, and is cancelling as easy as subscribing? Would we be comfortable seeing this output shared publicly with our name on it? Summary Positioning a consumer AI product as reflection and entertainment is a product decision that has to run through every layer, not a disclaimer at the bottom of the page. At ZodiVela, the structured parts of a reading are calculated or drawn deterministically and paired with curated meanings; AI writes the reflective interpretation within rules that forbid predictions, instructions and advice. Safety hooks detect health, legal, financial and crisis topics and change what the product does, with crisis signals stopping the reading entirely. Delight comes from craft, ritual, personal relevance and depth on demand, not from fear or urgency, and pricing through credits and subscriptions is kept clear and easy to cancel. Honest positioning builds the trust that makes a consumer brand last. ## From free test to paid report: designing an honest consumer funnel URL: https://oryvelon.com/insights/from-free-test-to-paid-report/ Category: Marketplaces & Consumer Products A free test followed by a paid report is one of the oldest patterns in consumer products: a personality quiz leads to a full profile, a credit check leads to a detailed report, a sample chapter leads to the book. It works because it lets people judge quality before they pay. It also has a bad reputation, earned by products that give away a teaser and hold the actual answer hostage. "You scored 72% — pay to find out what it means." Two Oryvelon companies use a version of the free test to paid report funnel: KeşifAtlası, for visa and relocation eligibility, and ZodiVela, for personal discovery readings. This article explains how we design the funnel so it is honest: what we give away, what we charge for, how we earn trust before asking for money, and how refunds work when an AI-assisted report falls short. Why this model suits some consumer AI products Not every product should use a free-test funnel. It suits products with three characteristics. The value is a one-time or occasional answer. A person planning a move abroad needs a clear assessment now, not a monthly subscription. Someone curious about their birth chart may want one detailed reading, then return months later. Quality is hard to judge in advance. People have been burned by generic answers in both categories. A free result lets them see the quality of the thinking before paying. The paid product has real extra depth. If the paid report is just the free result with more words, the model collapses. There must be genuinely more to deliver. Products with continuous value — a daily operations assistant like MerchNivo — suit subscriptions better. See Designing a recurring revenue loop. The honesty rule: the free result must stand on its own The single most important design rule we apply is this: the free result must be useful even if the person never pays. For KeşifAtlası, that means the free test returns the headline eligibility outcome — likely eligible, not eligible on current answers, or needs review — and the main reason. If someone fails a funding requirement, the free result says so. We do not hide the reason behind the paywall. For ZodiVela, it means the free reading is a real reading: a short reflection on the person's sun, moon and rising signs, or a single-card draw with a proper interpretation. Not a blurred preview. This can feel commercially risky. Why would anyone pay if the free result already answers the question? In practice, people pay for what the free result makes them want next: detail, structure, sources, a plan, a deeper exploration. The free result creates trust; trust creates the sale. A teaser creates suspicion, and suspicion creates refund requests. What goes in the free tier and what goes in the paid tier Drawing the line is a product decision we write down in each company's source of truth. A useful way to think about it: Free test Paid report Purpose Answer the headline question Help the person act on it KeşifAtlası Eligibility outcome per route, main reason Requirement-by-requirement breakdown, official sources, document checklist, timeline, what could change the outcome, comparison of suitable routes ZodiVela Short reading on core placements or a single card Full chart reading by life theme, multi-card spread, personal reflection prompts, a keepsake format Depth One screen Several sections, downloadable Personalisation Based on test answers Based on test answers plus optional detail the person adds Human route Link to ask a question Priority route for questions (where offered) The paid report is always built from the same underlying data as the free result. For KeşifAtlası, that means the same verified rules database and the same rule-set version; the report simply shows more of what the rules engine already knows. See Rules engines for eligibility. A paid report must never reach a different conclusion from the free result for the same answers. If it did, one of them would be wrong. Earning trust before the payment step People decide whether to pay in the few seconds after they see their free result. What they see then should answer the questions they are silently asking. What exactly will I get? Show a clear list of report sections, and where possible a real sample report for a fictional profile. A table of contents converts better than a promise, because it is specific. How much, all in? Show the price in the person's currency, including tax where applicable. No surprise fees at checkout. When will I get it? Most reports generate in seconds or minutes. Say so. If a report includes a human review step, give an honest time range. Who is behind this? A clear company name, contact route, and privacy notice. For KeşifAtlası, a visible note on when the rules for that route were last updated. The light "An Oryvelon company" endorsement helps here without turning the product into a group brand. What if it's not right? A plain-language refund policy linked right next to the payment button. What is this not? For KeşifAtlası: an informational assessment, not legal advice, and not a guarantee of any visa decision. For ZodiVela: a reading for reflection and entertainment, not a prediction or professional advice. Saying this before payment prevents misunderstanding afterward. See Positioning consumer AI as reflection and entertainment. The broader principle — setting correct expectations early so the AI output is judged against the right standard — is covered in Onboarding and trust in AI products. How paid reports are generated A paid report is where AI is most visible, and where quality control matters most. Our pattern is the same for both companies, with different sources of truth. Assemble verified inputs. For KeşifAtlası: confirmed answers, rule-set version, every rule evaluated and its result, sources and plain-language rule text. For ZodiVela: calculated chart placements or drawn cards, and the curated meanings for each. Generate section by section. Each report section has its own prompt and its own output schema. Generating in sections keeps each piece focused and makes failures easier to catch and retry. Validate every section. Outputs are checked against their schema. For KeşifAtlası, every requirement mentioned must map to a rule that was evaluated; nothing may be added. For ZodiVela, outputs are checked for predictive language, instructions and sensitive topics. See Structured outputs and schema validation. Fall back when needed. If a section fails validation after retries, it is replaced with a well-written templated version built from the structured data, rather than shipping a flawed paragraph. See AI fallbacks and degraded modes. Record versions. Each report stores the rule-set version, prompt versions and model route used, so it can be reproduced or investigated later. All of this runs through the group's AI gateway, which handles model routing, budgets, caching and logging, while each company keeps its own prompts, data and users. Pricing the report We are cautious about stating specific prices in general articles, because prices change and differ by market. The principles matter more than the figures. Price against the value of the decision, not the cost of generation. A relocation decision involves significant time and money; a clear, sourced report that saves hours of research is worth a meaningful fee. A reflective reading is a smaller, more personal purchase. Keep the cost of generation well below the price. Section-by-section generation, caching of curated content and routing simpler sections to smaller models keep the gross margin healthy. See Cost discipline for AI products. Offer one clear primary option. Too many tiers at the moment of decision slow people down. KeşifAtlası can offer one standard report per route; ZodiVela can offer a small number of report types clearly distinguished by content, plus credits or a subscription for regular users. Avoid dark patterns. No fake discounts from an inflated "original" price, no countdown timers, no pre-ticked add-ons. Refunds: a policy and a feedback loop AI-assisted reports can occasionally be wrong or poor. A report may have a section that failed silently, contradict the person's answers, or — in KeşifAtlası's case — reflect a rule that changed between test and purchase. A refund policy should handle these fairly and quickly. A policy that works for both companies: Clear grounds. A report is refundable if it is missing promised sections, contradicts the person's confirmed answers, contains a factual error against our own rules database, or was never delivered. Simple request. One form or email, no argument required for clear cases. Quick decision. A stated response time, kept. No refund for disagreement with a correct outcome. If the rules say a person does not currently meet a requirement, a correct report saying so is not a defect. We explain this before purchase so it is not a surprise afterward. Consumer rights respected. Local consumer law sets minimums in each market; our policy should meet or exceed them. This is a description of practice, not legal advice. Every refund reason is tagged and reviewed. Refunds are one of the most honest signals a paid product gets. A cluster of "section missing" refunds points to a generation or validation bug; "didn't match my situation" points to the questionnaire; "expected a guarantee" points to pre-purchase copy. Refund rate per report type is one of the metrics we look at in 30/60/90-day checkpoints. Worked example: KeşifAtlası Imagine a graduate in Türkiye considering a skilled-work route in another country. They take the free test: eight to twelve short questions depending on their answers. The result says they look likely eligible for one route and not eligible on current answers for another, because they do not yet hold a job offer. Both reasons are shown in plain language, with the date rules were last updated. Below the result they see what the paid report contains: a requirement-by-requirement breakdown for both routes, links to official sources, a document preparation checklist, a realistic timeline, what would change the second outcome, and a clear note that the report is informational and that authorities make visa decisions. They see a sample report for a fictional profile, the price in lira, and the refund policy. If they buy, the report is generated from the same answers and rule-set version, section by section, validated and delivered as a web page and a download. If they have a question the rules cannot answer — say, about a previous refusal — the report offers a route to a person rather than a guess. See Human escalation in AI products. Worked example: ZodiVela Imagine someone who enters their birth date, time and place. The free reading gives a warm, well-written reflection on their sun, moon and rising signs, with a journalling prompt. It is complete in itself. Below it they see the full chart reading: sections on relationships, work and creativity, personal growth and the year's themes, all written as reflection, each with prompts. There is a sample, a clear price, and a note that readings are for reflection and entertainment. Regular users can choose credits or a subscription instead of single purchases. If they buy and a section arrives visibly broken — say, repeating another section — the refund is immediate, and the defect is logged against the prompt version that produced it. Data in the funnel A free test collects personal information from people who have not yet decided to become customers. That calls for care. Ask only what the result needs. For KeşifAtlası, every question maps to a rule. For ZodiVela, birth time is optional and explained. Separate test data from marketing consent. Taking a free test is not consent to marketing. If we want to email someone later, we ask separately and clearly. See Marketing consent vs product consent. Set retention for abandoned tests. Answers from people who never create an account or buy should not sit indefinitely. See Data retention and deletion. Measure the funnel without tracking people across the web. Conversion from test to report can be measured with first-party, aggregate analytics. When the rules change between test and purchase KeşifAtlası has a problem most report businesses do not: the thing being reported on can change. A person might take the free test on Monday, come back on Friday to buy, and find that a rule for their route changed on Wednesday. We handle this explicitly rather than hoping it is rare. When a person returns to buy, the product checks whether the rule-set version for their route has changed since their test. If it has, they are told before payment, shown their updated free result, and asked to confirm. The paid report is always generated against the current rules for their intended timeline, and the report states which version it used. The same logic applies after purchase. If a rule changes shortly after someone buys a report, the report does not silently change, because it records the version that produced it. What we can do is identify affected reports and, where the person has agreed to service messages about their report, let them know an update is available. That is a product decision with a consent question attached, not an automatic marketing email. What we measure in the funnel A funnel like this is easy to measure badly. Test completions and purchases alone tell you very little. The metrics we find useful: Test completion rate by question step, to find where people give up. Free result to report page — the share of people who look at what the paid report contains. Report page to purchase, split by outcome. People with a "needs review" outcome behave differently from those with a clear result. Refund rate and reasons per report type and per prompt version. Validation failure and fallback rate per report section, which predicts refunds before they happen. Return visits within a few months, especially for ZodiVela, where a good first report often leads to a second. None of these require tracking individuals across other sites. Keeping measurement first-party and aggregate is a choice that fits the sensitivity of both products. Common mistakes Paywalling the answer. A free test that withholds the result destroys trust and invites complaints. A paid report that disagrees with the free result. If they are built from different logic, one of them is wrong. Build both from the same source. Vague report contents. "Get your full personalised report" is not a description. List the sections. Hiding the refund policy. People who can see a fair policy are more willing to buy, not less. Ignoring refund reasons. Each refund is a free bug report. Retargeting everyone who took the test. It feels invasive, and for sensitive categories like relocation it can be. Summary A free test to paid report funnel works when the line between free and paid is honest. The free result must answer the headline question on its own; the paid report adds depth, structure, sources and next steps, built from the same verified data so the two can never disagree. At KeşifAtlası that data is a versioned rules database; at ZodiVela it is calculated chart placements or drawn cards paired with curated meanings. Trust is earned before payment with a clear list of contents, a sample report, the full price, delivery time, what the report is not, and a visible refund policy. Reports are generated section by section, validated and versioned, with fallbacks for failures, and every refund reason feeds back into the product. ## Why a services company lives inside a company builder URL: https://oryvelon.com/insights/services-company-inside-a-group/ Category: Marketplaces & Consumer Products People who look at Oryvelon for the first time often ask a fair question. If the group has a digital and advertising agency in it, isn't Oryvelon really an agency with some side products? It is not, and the reason is deliberate structure. Oryvelon is a digital company builder: it builds, owns and operates independent companies. It does not sell services to clients. Client work lives in WeAreMedia, the group's services company, which contracts, delivers and bills that work under its own name. This article explains why a services company inside a company builder makes sense, why we keep it separate from Oryvelon itself, where the boundaries sit, how pricing works when group companies are also clients, and what a services business teaches a portfolio of products. If you are a founder with an agency background thinking about building products — a very common path — the structure may be useful. What a company builder is, and is not A company builder creates companies and operates them for the long term. Its revenue ultimately comes from the companies it owns: their subscriptions, marketplace fees, reports and product sales. See What is a digital company builder?. An agency sells time and expertise to clients. Its revenue comes from retainers, projects and fees. Its product is delivery. Both are legitimate businesses. They have different economics, different risks, different sales motions and different obligations to customers. Mixing them in one entity creates confusion on every one of those dimensions: Company builder (Oryvelon) Services company (WeAreMedia) Earns from Companies it owns and operates Client projects and retainers Main asset Products, data rules, brands, infrastructure People, processes, reputation, client relationships Customer Each product's users Brands that hire the agency Main risk Product-market fit, capital allocation Utilisation, client concentration, delivery quality Data held Each company's own users, separately Client data, per engagement Success looks like Companies that can stand alone Clients who stay and refer That table is why "is Oryvelon an agency?" gets a clear no. The agency work is real and valuable. It just belongs somewhere else. Why keep services inside the group at all If the models are so different, why not run WeAreMedia as an entirely unrelated business? Several reasons. It already existed and works. The founders' experience in e-commerce, Shopify and performance marketing is where much of Oryvelon's thinking comes from. A services company is the natural home for that expertise when it is applied to other people's brands. It keeps the group close to real problems. Product companies can drift into building for imagined customers. A services team spends its days inside real businesses: stores with messy catalogues, ad accounts with broken tracking, founders who need a website by next month. That exposure is hard to get any other way. It is a capable delivery partner for group companies. New companies need websites, launch campaigns, creative and analytics setup. WeAreMedia can deliver those — as a supplier, under the same terms as any client. It is a business in its own right. WeAreMedia has its own revenue, costs and clients. It is held to the same standard as every other company in the portfolio: it must be able to stand alone. See Designing every company so it could stand alone. Boundary 1: the brand Oryvelon runs a house of brands. Each company has its own name, domain and voice, with a light "An Oryvelon company" endorsement. See Naming, domains and brand architecture. For services, that means: Services are sold as WeAreMedia, on wearemedia.us, with WeAreMedia proposals, contracts and invoices. Oryvelon's site does not offer services. If someone contacts Oryvelon asking for a website or ad campaign, we point them to WeAreMedia. The contact page says the same. WeAreMedia's portfolio shows client work, not Oryvelon's own companies presented as if they were clients, unless they genuinely were and the work is accurately described. This is not about hiding the relationship. The endorsement is visible. It is about making sure a person always knows which company they are dealing with, and which company is responsible. Boundary 2: client data This is the boundary we are strictest about, because it is where a services company inside a product group could go most wrong. A services team often has deep access to client systems: Shopify admin, ad accounts, analytics properties, customer email lists, sometimes sales data. It would be technically easy — and a serious breach of trust — to let that data flow into product companies. Imagine a product team quietly training a feature on patterns from agency clients' stores. Our rules: Client data stays inside each client engagement. It is accessed through the client's own accounts, under the client's permission, for the client's work. No client data reaches any Oryvelon product. Not MerchNivo, not analytics, not AI prompts, not training or evaluation sets. MerchNivo works only with the data of stores that connect to it directly, under its own terms. Access is granted per engagement and removed when it ends. Named individuals, the least access needed, two-factor authentication everywhere. See Least-privilege access for small teams. Credentials stay out of chat and documents. Client passwords and API keys are handled through proper secrets practices. AI tools used in client work respect the same boundary. If WeAreMedia uses AI to draft ad copy or analyse a campaign, that work runs within WeAreMedia's own setup and the client's permissions, not inside a product company's prompts or knowledge. These rules are an application of a group-wide principle: shared infrastructure, separate data. Services are the clearest test of it. Boundary 3: what "learning from services" is allowed to mean We do say that WeAreMedia's work informs the group. It is worth being precise about what that means, because it would be easy to overclaim. What flows from services to products is general operating knowledge. For example: "store owners consistently struggle to notice when a best-selling product is about to go out of stock"; "refund spikes after a promotion are often a sizing or description problem, not a product problem"; "small brands rarely have clean UTM tagging". These are the kinds of observations any experienced operator carries in their head. What never flows is client data or client-identifying detail. No store names, no figures, no customer lists, no screenshots of dashboards. A useful test: could this insight have been written by an experienced e-commerce consultant from general experience, without reference to any particular client? If yes, it can inform a product. If it depends on a specific client's data, it stays in that engagement. That is also why Noveniq, our own DTC technology accessories store, matters. When MerchNivo needs to test a feature on real store data, it has a real store the group owns and operates, with data the group is entitled to use. See Using a real store as a testbed. Boundary 4: pricing and money between companies When group companies use WeAreMedia, the easy mistake is to do the work "for free because we're all one group". We avoid that. Group companies are clients. Work is scoped, estimated and invoiced like any other engagement. Terms can reflect the relationship, but they are written down. Why it matters: - A product company that gets free marketing looks healthier than it is. When it is assessed at its 30/60/90-day checkpoint, its true costs must be visible. - A services company that gives away time looks less profitable than it is and loses capacity for paying clients. - If a company is ever sold, spun out or partnered, clean arrangements make it separable. Capacity is a second issue. If WeAreMedia's best people are always on group work, client delivery suffers. Group work is planned into capacity like any other client, not squeezed in. See Unit economics per product. Boundary 5: people and roles A small group shares people, and pretending otherwise would be dishonest. The founders are involved across companies; some specialists work on both services and product work. What we separate is roles and access, not necessarily people: When someone works on a client engagement, they act for WeAreMedia and use WeAreMedia's access to that client's systems. When the same person works on a product, they use that product's access only. Access is per company and per engagement, reviewed regularly. This is less about bureaucracy than about making the right behaviour the easy behaviour. If a person's product account cannot see client systems, a mistake is much less likely. How WeAreMedia fits the group's operating model WeAreMedia follows the same operating standards as the rest of the portfolio, adapted to a services business: A source-of-truth document defining what WeAreMedia offers — web and Shopify development, social media, paid media, SEO, creative, content and AI automation — who it serves, how it earns, and what it will not do. See Source-of-truth documents. Its own metrics. Utilisation, client retention, margin per engagement and pipeline, rather than product metrics like activation or churn. Its own brand, domain, email and analytics, with separate tracking standards. Regular review in the group's operating cadence, alongside product companies. Worked example: a request that arrives at the wrong door Imagine a fashion brand emails Oryvelon: they have seen the group's work and want help rebuilding their Shopify store and running a launch campaign. Oryvelon replies that the group does not offer services directly, and introduces WeAreMedia. From that point, the conversation, proposal, contract and invoice are all WeAreMedia's. The brand's store access is granted to named WeAreMedia team members for the duration of the project. Six months later, the brand sees MerchNivo and wants to try it. That is a completely separate relationship. The brand signs up to MerchNivo directly and connects its store under MerchNivo's terms. MerchNivo does not inherit anything from the WeAreMedia engagement — not the brand's data, not its account history, not its contacts. If the brand wants to stop using either service, each relationship ends independently. It is slightly more administration than a single "Oryvelon does everything" relationship. It is also exactly what a careful client would want. What a services business teaches a product portfolio Beyond specific observations about stores and campaigns, running a services company alongside products teaches habits that product teams tend to lack. Selling is a skill, not a launch event. An agency has to win every client through conversation, proposals and trust. Product companies sometimes expect a good landing page to do all that work. Watching how services are actually sold — what questions prospects ask, what reassures them, what makes them hesitate — shapes how our product sites are written and how onboarding explains things. See Onboarding and trust in AI products. Deadlines are real. Client work has launch dates that cannot move. That discipline carries into product work: a phase-zero site shipped this month beats a perfect one next quarter. Operators care about outcomes, not features. A store owner does not want "an AI assistant". They want to know which products to reorder, why refunds jumped and whether the new campaign is working. Every product brief in the group is better for having been tested against that kind of plain question. Reporting has to be honest. Agencies that inflate results lose clients eventually. The same attitude — show the numbers as they are, from the system that owns them — is built into MerchNivo's approach of taking figures from the store rather than from a model. When this structure is the wrong choice We think the structure works for us. It is not right for everyone, and it has costs. It adds administration. Separate contracts, separate invoices, separate access and separate reporting take time. A two-person team with one product and a few clients may reasonably decide it is not worth it yet. It can split attention. Founders involved in both services and products are always choosing where to spend the next hour. Without clear priorities and protected product time, services win by default because clients are waiting. It can confuse the market if the boundaries are weak. If the websites, email addresses and sales conversations blur the two, the structure adds complexity without adding clarity. It only works if the services business is healthy. A struggling agency inside a group becomes a drain on attention and capital. WeAreMedia is judged on its own results, like any other company. If you are considering a similar setup, a short test: could you explain to a client in one sentence which company they are contracting with, who has access to their data and what happens to that data when the project ends? If the answer takes a paragraph, the boundaries need work before the structure will help. A checklist for structuring services inside a group For founders with an agency who are starting to build products, these are the decisions we would write down early: Which legal entity and brand contracts client work, and which never does? Where do inbound service requests go if they arrive at the group or a product company? What client data does the services team hold, where, and for how long after an engagement ends? Which product companies are explicitly forbidden from using client data, and how is that enforced in access, not just policy? How are group companies charged when they use the services company? What metrics judge the services business, separately from product metrics? Who reviews access across companies and engagements, and how often? Common mistakes founders make when mixing services and products Letting the agency name become the group name. Once clients know you as an agency, every product you launch looks like a side project. Separate brands avoid that. Treating client data as a product asset. Even if contracts technically allow some use, the trust cost is high and the practice rarely survives scrutiny. Subsidising products with unbilled services. It hides the true cost of the product and starves the services business. Letting services revenue decide product strategy. Services pay monthly; products take time. It is easy to keep pushing product work back because a client needs something now. Protected time and separate goals help. Building "productised services" and calling them software. Some services can become products, and that is a legitimate path. But a product that needs a human behind every delivery is still a service, and it should be priced and staffed as one. Summary Oryvelon is a company builder, not an agency. Client work — web, Shopify, paid media, SEO, creative, content and AI automation — is contracted, delivered and billed by WeAreMedia, a separate services company within the group with its own brand, domain, metrics and source of truth. The boundaries are explicit: services are sold only as WeAreMedia; client data stays within each engagement and never reaches Oryvelon products; group companies that use WeAreMedia are scoped and invoiced as clients; and people's access is separated by company and engagement. What flows from services into products is general operating knowledge, never client data, and real-store testing happens on Noveniq, which the group owns. The structure keeps each business honest about its economics and gives clients the clarity they deserve. ## Services, education and digital products: building a founder-led expertise business URL: https://oryvelon.com/insights/services-education-and-digital-products/ Category: Marketplaces & Consumer Products Some of the most durable small businesses are built on one person's craft. A skilled professional builds a reputation delivering a service, other people start asking to learn it, and eventually the know-how gets packaged into something people can buy without booking an hour of the founder's time. That arc — services, education and digital products — is the model behind Sinem Keser Beauty Academy, the beauty and education company in the Oryvelon group. It offers professional beauty services by appointment, structured beauty education, and digital learning products, all under a founder-led brand. This article describes how we think about a founder-led expertise business: what each revenue line is for, how they connect, what to build first, how to protect the founder's time, and how to handle data and brand when one name sits on all three. The pattern applies well beyond beauty — to photographers, chefs, coaches, designers, fitness professionals and many other crafts. Three revenue lines, three different businesses It is tempting to think of services, courses and digital products as one business with three price points. They are closer to three businesses that share a founder and a brand. Services Education Digital products Customer Someone who wants the result Someone who wants to learn the craft Someone who wants guidance or tools at their own pace What they buy An appointment A course, workshop or programme A guide, video lesson, template or toolkit Delivery In person, founder or trained team Live, in person or online, with the founder as teacher Self-serve, available any time Scales with Hours and chairs Cohort size and course design Distribution Main constraint Founder time and location Teaching quality and support Discovery and trust Main risk Burnout, no-shows Poor outcomes for students Low perceived value, piracy Revenue shape Steady, capped Lumpy, around course dates Low per sale, potentially repeating Each line needs its own offer, pricing, funnel and success measures. They share a reputation, which is the most valuable asset in the business and the one most easily damaged. Services: the foundation of credibility The service is where expertise is proven every day. Clients see the result; the founder refines technique; the portfolio grows. Without this, education and digital products are just claims. What makes a services line work in a founder-led business: Frictionless booking. People should be able to see services, prices, durations and availability and book in a few taps. Phone-and-message booking does not scale and loses people outside working hours. Clear policies. Deposits, cancellations and lateness, stated up front. No-shows are one of the biggest hidden costs in appointment businesses. Aftercare and follow-up. Clear care instructions and a well-timed reminder for the next appointment turn a single visit into a relationship. A portfolio with consent. Before-and-after photos are powerful marketing. They must only be used with the client's explicit, recorded permission, and people must be able to withdraw it. The ceiling on services is obvious: the founder has a fixed number of hours. A trained team can extend that, but only if quality holds. That ceiling is exactly why the other two lines matter. Education: turning craft into teaching Many clients, and many aspiring professionals, eventually ask the same question: can you teach me? Education turns the founder's expertise into structured learning. It is a different skill from doing the work. A great practitioner does not automatically become a great teacher, and courses deserve the same design care as any product. Principles we apply to the academy's education line: Define the outcome. "After this course you will be able to..." in concrete, honest terms. No promises of income or guaranteed careers. Structure before content. Modules, sequence, practice, assessment. A course is a path, not a pile of videos. Practice and feedback. Beauty is a hands-on craft. Courses need supervised practice, feedback on students' own work and clear standards. Right format for the goal. Intensive in-person workshops for hands-on technique; online modules for theory, product knowledge and business skills; a combination for longer programmes. Clear certificates. If a course awards a certificate, say plainly what it represents — completion of the academy's course — and what it does not, such as a state licence where one is required. Education changes the founder's economics. A course taught once to a group earns more per founder hour than individual appointments. A well-designed online module can be taught once and taken many times. Digital products: expertise without the appointment Digital products are the most scalable line, and the one where quality is hardest to prove before purchase. In a beauty academy, digital products can include recorded technique lessons, product and tool guides, step-by-step routines, checklists for professionals setting up their own practice, and templates for pricing, client records or consultation forms. Each should solve one clear problem for one clear audience. What we think makes digital products succeed: Specificity. A focused guide on one technique beats a general "everything about beauty" bundle. Specific products are easier to explain, price and trust. Visible quality. Free samples — a short lesson, a page from the guide — let people judge before paying. This is the same principle we apply to free tests and paid reports. Clear format and access. What exactly is included, how it is delivered, whether access expires, which devices it works on. Honest claims. Beauty is a category prone to overpromising. Products should describe what they teach, not promise transformations. Digital products also have a strategic role: they are the entry point for people who are not near the founder's location, not ready to book a course, or simply curious. A good low-priced product can be someone's first purchase from the brand. Recurring loops: how the three lines feed each other Individually, each line has limits. Together, they form loops that make the business more stable. We describe the general idea in Designing a recurring revenue loop; here is how it looks in an expertise business. The client loop. A person books a service, has a good experience, receives aftercare and a timely reminder, and books again. Repeat appointments are the simplest recurring revenue in the business. The client-to-student loop. Some clients become curious about the craft itself. A client who has seen the founder's work up close is an unusually well-informed prospective student. The student-to-customer loop. Students need tools, templates and continuing education after their course. Digital products and advanced courses serve them. A graduate who starts practising professionally may keep buying for years. The content loop. The founder's work, teaching and advice, shared as content, brings new people to the brand. Some book services; some take courses; some buy a digital product. Content is how the other loops get fresh people. The alumni loop. Graduates who do well become the academy's strongest proof. With their permission, their work and stories bring in the next cohort. None of these loops works if a customer feels sold to at every turn. They work when each step genuinely serves the person at the moment they are ready for it. Protecting the founder's time In a founder-led business, the founder's hours are the constraint on everything. A useful long-term goal is to move a growing share of revenue from time-bound to time-independent sources, without losing the hands-on practice that makes the expertise real. Practical steps: Price services to reflect scarcity. If the founder's calendar is full weeks ahead, prices are probably too low. Some services can be delivered by trained team members at a different price. Batch teaching. Run cohorts on set dates rather than one-to-one training on request. Record what repeats. If the founder explains the same theory in every course, record it once, well, and use class time for practice. Automate administration. Booking, reminders, deposits, course enrolment, digital delivery and receipts should not need the founder. Protect creation time. Digital products do not get made in the gaps between appointments. They need scheduled time. The aim is not for the founder to stop practising. It is for practice to be a choice, not the only way the business earns. One brand, two sites, clear roles A founder-led business usually carries the founder's name, which raises a brand question: where does the person end and the business begin? In our group, Sinem Keser has a personal authority site, sinemkeser.com.tr, which presents her as a professional and founder. The academy lives at sinemkeser.com, which is where services are booked, courses are listed and digital products are sold. This mirrors how we think about founder sites across the group: the person's authority and the company's offers are related but separate. See Founder authority sites. The separation is practical: The authority site carries biography, philosophy, media, speaking and the founder's broader work, including her role as co-founder of Oryvelon. The academy site carries offers, booking, course details, policies, product pages and support. Each has its own analytics and clear links to the other, tagged consistently. See UTM standards across brands. Both are written so that search engines and AI assistants can understand what each is. Clear structured data describing the person, the business and its offers helps people find the right site for their question. See Search and AI visibility for companies. Data: clients, students and customers are different An expertise business collects data from three kinds of people for three different purposes, and treating them all as "the mailing list" is a common mistake. Clients share contact details, appointment history and sometimes information relevant to their treatment, such as sensitivities a professional needs to know about. That information is for safe service delivery and should be accessed only by the people providing it. Students share enrolment details, progress and their own work for feedback. That is for teaching. Digital product customers share purchase and access details. That is for delivery and support. Consent to marketing is separate from all of these. Booking an appointment is not consent to course promotions; buying a guide is not consent to service offers. Ask clearly, record the answer, and make it easy to change. See Marketing consent vs product consent. Retention matters too. Old appointment notes, lapsed student records and abandoned enrolment forms should not be kept forever. See Data retention and deletion. And none of this data is shared with other Oryvelon companies, in line with Shared infrastructure, separate data. Where AI helps, and where it does not AI has a modest, useful role in a founder-led expertise business. It does not replace the expertise. Useful: drafting course outlines from the founder's notes for her to edit; turning a recorded lesson into a written summary; answering routine questions about booking, policies and course logistics from an approved knowledge base; helping write product descriptions that the founder then checks. Not appropriate: giving personalised treatment advice, assessing skin or other conditions, or grading practical work without the founder or a trained teacher. Those require a professional's judgment. Where an assistant is used on the site, it should stay within approved information and hand questions it cannot answer to a person. See Human escalation in AI products. This is the same rule we apply across the group: AI can explain and assist; verified knowledge and qualified people decide. Worked example: a year in the life of one customer Imagine someone who finds the academy through a short technique video. They buy a low-priced digital guide on the same technique. They like the clarity of the teaching. A few months later, they book a service at the academy. The booking is simple, the policy is clear, aftercare arrives the same day and a reminder arrives when it is time to rebook. They come back twice more. Along the way, they decide they want to learn the craft properly. Because they already know the founder's standards, they enrol in an in-person course with confidence. After graduating, they buy templates for setting up their own practice and later an advanced online module. Each step was optional. Each served them when they were ready. And each came through a clear, honest offer rather than pressure. That is what a recurring loop looks like from the customer's side. What to build first For a founder starting this kind of business, the order matters. Services first. They prove the expertise, generate income and create the portfolio and reputation everything else depends on. One course next. Start with the most-requested topic, taught in person or live, so feedback is direct. Refine it before recording anything. Digital products from what repeats. The questions students and clients ask most often become the first guides and lessons. Systems throughout. Booking, payments, enrolment and delivery should be automated early; they are cheap to set up and expensive to retrofit. At each stage, the business should be able to stand alone. See Designing every company so it could stand alone. Measuring each line separately Because the three lines are different businesses, a single revenue figure hides most of what matters. We look at each line on its own terms. Services: utilisation of available appointment hours, no-show and late-cancellation rate, rebooking rate within the typical interval for each service, and revenue per founder hour compared with revenue per team-member hour. Education: enrolment against capacity for each course date, completion rate, student feedback on specific modules, and — with graduates' permission — whether they go on to practise. A course that fills but leaves students unable to do the work is a failing course, however good the revenue looks. Digital products: conversion from free sample to purchase, refund rate and reasons, and how many buyers go on to a second purchase or a course. Across the loops: the share of new students who were previously clients or product buyers, and the share of revenue that does not depend on the founder's hours that month. The second number is the clearest single signal of whether the business is becoming less fragile over time. These measures feed the same continue, stop or scale reviews we run across the portfolio. A course that consistently underfills or a product that attracts refunds is reworked or retired, the same way a product feature would be. Common mistakes Launching courses before the reputation exists. Education sells on proven expertise. Recording everything at once. A large library of unrefined videos is harder to sell than one excellent course. Overpromising outcomes. Income claims and "transformation" language erode trust and can conflict with consumer protection rules. Pooling all contacts into one list. Clients, students and customers have different needs and different consents. Letting the founder remain the only delivery channel. Without team training, recorded content and automation, growth stops at the founder's calendar. Summary A founder-led expertise business can earn from the same craft in three ways: delivering the service, teaching it and packaging it into digital products. At Sinem Keser Beauty Academy, appointments build credibility and steady income, courses turn expertise into structured teaching, and digital products make it available to people anywhere, at their own pace. Each line has its own customer, offer, price and risks, and they are connected by recurring loops: clients rebook and some become students, students buy tools and advanced courses, content and alumni bring in new people. The founder's time is protected by moving steadily toward time-independent revenue, with separate authority and academy sites, careful handling of client, student and customer data, and AI kept to assisting rather than advising. ## Search and AI visibility for companies: how to be the clear, consistent answer URL: https://oryvelon.com/insights/search-and-ai-visibility-for-companies/ Category: Growth, Search & Measurement Search and AI visibility for companies used to mean one thing: ranking in Google. Today a growing share of people ask an AI assistant first — "what is a good tool for Shopify store operations?", "is there a service that checks visa eligibility for Türkiye?", "who runs CastLyra?" — and the assistant answers from what it has read, retrieved or been told. The company that appears in that answer, described correctly, is the company that gets considered. We run a group of independent companies, each with its own brand and domain. That makes discoverability both easier and harder. Easier, because shared standards mean every company starts with a sound technical base. Harder, because a group of related brands can easily confuse machines: which name belongs to which product, which site is authoritative, who is the parent and who is the child. This note explains how we think about being discoverable in search engines and AI assistants, what we actually do across Oryvelon's companies, and what we deliberately refuse to do. Why search and AI visibility now overlap It is tempting to treat "AI visibility" as a new discipline with new tricks. In practice, the foundations are shared. Search engines build an understanding of entities — organisations, people, products, places — and the relationships between them. They reward pages that clearly answer a question and come from a source that appears trustworthy on that topic. AI assistants do something similar through a different route. Some answer from what was in their training data. Many now search the web at answer time, retrieve a handful of pages, and summarise them. Either way, the assistant is looking for text that states clearly what something is. So both systems reward the same things: An unambiguous entity. One name, one domain, one description, used consistently. Crawlable, readable pages. Content that loads without heavy scripts, with sensible headings and plain HTML. Clear answers. Pages that say directly what a product does, who it is for and what it does not do. Corroboration. The same facts appearing in several places the machine already trusts. The differences are at the edges. AI assistants seem especially sensitive to quotable, self-contained statements — a sentence that can be lifted into an answer without losing its meaning. And because they compress everything into a few lines, ambiguity costs more. If your homepage says one thing and your social profiles say another, an assistant may confidently repeat the wrong one. Entity consistency: the foundation most companies skip An entity is a thing a machine can recognise and keep separate from other things. For a company, entity consistency means that everywhere the company appears — its own site, its structured data, its social profiles, directories, press mentions, the parent group's site — the core facts agree. We keep a short, fixed set of facts for each company in its source-of-truth document: Fact Example: KeşifAtlası Name (exact spelling) KeşifAtlası Primary domain kesifatlasi.com One-sentence description Visa and relocation eligibility grounded in a verified rules database Category AI product What it is not Not legal advice; the AI explains rules, it does not decide eligibility Parent relationship An Oryvelon company That table becomes the input for everything else: page titles, meta descriptions, the Organization markup on the company's own site, the entry on the group site, the bio on social profiles, and what we say to journalists. When the facts change, they change in the source of truth first and flow outward from there. This sounds bureaucratic. It saves a surprising amount of trouble. A common failure in multi-brand groups is description drift: the product page says "AI assistant for merchants", the social bio says "Shopify analytics", a directory listing says "e-commerce agency", and a press article says something else again. A human might reconcile these. A machine often cannot, and it may merge, split or mislabel the entity as a result. Names with special characters Some of our names carry non-ASCII letters. KeşifAtlası is the obvious case. The domain is kesifatlasi.com because domains are simpler that way, but the brand name keeps its Turkish characters. We are consistent about which form appears where: the brand name with diacritics in text and markup, the ASCII form only in URLs and handles where the platform requires it. We also make sure the plain-ASCII spelling appears somewhere natural on the site, so a person or a model searching for "Kesif Atlasi" still lands in the right place. Parent and child, stated once and clearly The relationship between the group and its companies is a frequent source of confusion. Is WeAreMedia part of Oryvelon? Is Oryvelon an agency? Is MerchNivo a feature of something bigger? We answer those questions explicitly, in plain text, in the places machines read. The Oryvelon site says that Oryvelon is a digital company builder, that it does not sell agency services, and that WeAreMedia is the group's services company. Each company page on oryvelon.com links to the company's own domain. The companies use a light "An Oryvelon company" endorsement rather than folding the group name into their product name. The reasoning behind that structure is in Naming, domains and brand architecture. The goal is for an assistant asked "is Oryvelon an agency?" to find a direct, quotable sentence that answers "no", and to find it in more than one place. Helpful content: answer the questions customers actually ask Search engines have spent years trying to demote pages written for ranking rather than for people. AI assistants inherit a version of the same preference: when they retrieve pages to answer a question, a page that answers it directly is far more useful than one that circles around a keyword. For us, helpful content starts from real questions. Each company keeps a running list of what prospective customers ask before buying, what they ask during onboarding and what support hears repeatedly. Those questions become pages, FAQ entries and guides. A few examples of what this looks like in practice: MerchNivo gets questions like "where do the numbers in my briefing come from?" The honest answer — from the store itself, never from the model — is also the answer that builds trust. It belongs on the site in plain words. See AI explains, verified data decides. KeşifAtlası gets questions about specific visa routes. Its content explains how rules work and what the free test checks, and is careful to say what the product cannot do: it does not give legal advice and routes hard cases to people. See Rules engines for eligibility. ZodiVela answers "is this a prediction?" directly: readings are framed for reflection and entertainment. Saying so clearly is both honest and good for discoverability, because it gives assistants an accurate one-line description. Depth beats volume One detailed page that fully answers a question tends to outperform ten thin pages that each answer a fragment. It also ages better. We would rather publish a small number of pages we are proud of and keep them current than build a large library that nobody maintains. This is also why these Insights articles are long. They are written to be the complete answer to a question a founder or operator might ask, not a teaser for a sales call. Honest limits are content too Pages that describe what a product does not do are unusually valuable. They prevent the wrong customers from signing up, reduce support load, and give AI assistants the context they need to describe a product accurately. When an assistant says "KeşifAtlası checks eligibility against a rules database but does not provide legal advice", that is a better outcome for everyone than a vaguer, more flattering summary. No programmatic spam Programmatic SEO — generating hundreds or thousands of pages from a template and a keyword list — is attractive because it is cheap and sometimes produces a burst of traffic. "Visa requirements for [nationality] moving to [country]" multiplied across every combination. "Best Shopify app for [niche]" across every niche. Now add a language model to write the filler and the cost per page falls close to zero. We do not do this for any company. The reasons are practical as much as principled. It creates risk that outlasts the traffic. Search engines explicitly target scaled, low-value content. A site that gets classified that way can lose visibility across all its pages, including the good ones. It is usually wrong in the details. For a product like KeşifAtlası, a templated page about a specific visa route that has not been checked against the verified rules database is worse than no page. It can mislead someone making a real decision. Our principle that the AI explains and verified data decides applies to marketing pages as much as to the product. It pollutes what AI assistants learn about you. If an assistant retrieves a thin, generic page from your site, that page becomes part of how it describes you. Mediocre pages dilute the clear description you worked to establish. It cannot be maintained. Rules change, products change, prices change. A thousand templated pages means a thousand pages that quietly go out of date. There is a legitimate version of structured, data-driven pages: when each page is backed by verified, specific data that genuinely differs between pages and is kept current. A store's product pages are an example. Noveniq's catalogue pages exist because each product is real, in stock or not, with its own specifications. The test we apply is simple. Would this page be useful to a person who landed on it with no other context? If the answer depends on luck, we do not publish it. Technical foundations every company inherits Because the group shares infrastructure, each company starts with the same technical baseline. None of it is exotic. All of it matters. Server-rendered or static HTML for public pages, so crawlers and AI retrieval tools see the content without executing heavy scripts. One canonical URL per page, with redirects from variants (with and without www, trailing slashes) handled once, at the edge. XML sitemaps listing only indexable pages, with honest last-modified dates. A robots.txt that allows reputable search and AI crawlers to read public pages, while blocking admin panels, form handlers and thank-you pages. Structured data for the organisation, articles, breadcrumbs and FAQs, matching what is visible on the page. An llms.txt file that gives AI systems a concise, curated map of the site and the most important facts. Fast pages with sensible images and minimal third-party scripts. The details of the machine-readable layer — llms.txt, schema.org markup, sitemaps and what oryvelon.com publishes — are in llms.txt and structured data. Phase-zero sites follow the same rules from day one. Even a company that is still being validated gets a clean, honest, crawlable page describing what it is, as described in Phase-zero sites. It is much easier to start consistent than to repair an inconsistent entity later. Each company earns its own visibility A group site could, in theory, try to rank for every query its companies care about. Oryvelon could publish pages about Shopify operations, visa eligibility, creative talent and astrology, and funnel traffic onward. We deliberately do not. Each company earns visibility on its own domain, for its own topics. The group site ranks for what the group is: a company builder, how it works, which companies it operates. It links to each company, and it writes about the operating principles behind them, but it does not try to become a second home for each company's customers. This follows from our view that every company should be able to stand alone. If MerchNivo's visibility depended on oryvelon.com, separating MerchNivo from the group would damage it. Search equity, like data, belongs to the company that earned it. It also produces clearer signals. When a person searches for a Shopify operations assistant, the best result is MerchNivo's own page, written for merchants. Not a group page written for investors and partners. Founders, authority and trust AI assistants and search engines both pay attention to who is behind a company. Our founders, Uğur Keser and Sinem Keser, each have their own authority site: ugurkeser.art and sinemkeser.com.tr. These sites describe their work and link to the group and to relevant companies in context. What they do not do is act as link farms. A founder's site that links to every company from every page, with keyword-rich anchor text, looks manipulative to machines and to people. We keep those links contextual: a founder writing about e-commerce operations can naturally mention MerchNivo; a page about brand and creator work can mention CastLyra. The full approach is in Founder authority sites. Measuring visibility without fooling ourselves AI visibility is hard to measure precisely. There is no universal report that tells you how often an assistant mentions your company. We use a mix of signals, and we are careful not to over-read any of them. Search console data for each company's domain shows queries, impressions and clicks. It is the most reliable signal we have for traditional search. Referral traffic from AI assistants appears in analytics when assistants link to a source and people click. It is a partial signal — many answers are read without a click — but a useful trend. Periodic manual checks. Once a quarter, for each company, someone asks a small set of fixed questions in several AI assistants and records whether the company appears and whether the description is accurate. "What does MerchNivo do?" "Is Oryvelon an agency?" "Which service checks relocation eligibility from Türkiye?" The point is not a score. It is to catch inaccurate descriptions early, trace them back to their source, and fix the source. Consistent campaign tagging so that traffic from our own links, newsletters and founder sites is not confused with organic discovery. See UTM standards across brands. All of this happens per company, in each company's own analytics property, without joining users across brands. See Analytics without surveillance. Common mistakes we see Across many years of working on e-commerce and digital businesses, the same visibility mistakes come up again and again. Different descriptions everywhere. The homepage, the social bio and the directory listing each describe a slightly different company. Heavy client-side rendering for key pages. The product description only exists after JavaScript runs, so some crawlers see an empty shell. Blocking AI crawlers by accident. A security plugin or a copied robots.txt blocks every non-mainstream bot, and the company then wonders why assistants describe it from outdated sources. Markup that disagrees with the page. Structured data claims ratings, prices or FAQs that are not visible to visitors. Thin location or variant pages. "Service in [city]" pages with nothing specific to each city. No statement of what the product is not. Assistants fill the gap with guesses, often flattering and often wrong. Treating the parent brand as a shortcut. Pointing every sub-brand's authority at the group site instead of building each company's own reputation. Each of these is cheap to avoid at the start and expensive to fix later. A practical checklist for a new company When a new company joins the group, its visibility setup follows a short checklist. The source-of-truth document contains the fixed entity facts: exact name, domain, one-sentence description, category, what it is not, and the parent relationship. The homepage states those facts in plain text near the top of the page. Organization structured data on the company's own domain matches the source of truth. The company's page on oryvelon.com uses the same description and links to the company's domain. Sitemap, robots.txt and llms.txt are published and checked. A first set of helpful pages answers the five to ten questions prospective customers ask most. A search console property and a dedicated analytics property exist for the domain, owned by the company. Social profiles use the same name and description. A small set of fixed questions is recorded for the quarterly AI-assistant check. None of it requires a large budget. It requires consistency and a willingness to write things down. Summary Search and AI visibility for companies rests on the same foundations: a consistent entity, helpful pages, clean technical basics and honest descriptions that machines can quote. We give every Oryvelon company a fixed set of facts in its source of truth and repeat them everywhere the company appears; we publish pages that answer real questions in depth, including what each product does not do; and we refuse programmatic spam because it trades short-lived traffic for long-lived risk and inaccurate descriptions. Each company earns visibility on its own domain, the group site explains the relationships clearly, and we measure per company without joining users across brands. ## llms.txt and structured data: the machine-readable layer of a company website URL: https://oryvelon.com/insights/llms-txt-and-structured-data/ Category: Growth, Search & Measurement A company website has two audiences. People read the pages. Machines — search engine crawlers, AI assistants, feed readers, link previews — read the pages too, but they read them differently. They look for signals that tell them, without guessing, what an organisation is, who runs it, what each page is about and which pages matter. The machine-readable layer is where those signals live: llms.txt, schema.org structured data, sitemaps, robots.txt and a few related files. None of them is magic. Together they remove ambiguity, which is most of what "being understood" means for a machine. This article explains what each piece does and what oryvelon.com publishes. We run a group of independent companies with their own domains, so our setup has one extra job: describing the relationship between a parent group and its companies clearly, without blurring them into one entity. The broader approach to discoverability is in Search and AI visibility for companies. What the machine-readable layer is for Before the details, it helps to be clear about the goal. Structured data and llms.txt do not make a weak page strong. They do three narrower, useful things. Identify entities. They state, in a fixed vocabulary, that this is an organisation with this name, this logo, these founders and these related organisations. Classify pages. They say that this page is an article, that one is an FAQ, this other one is a contact page. Point to what matters. They list the pages worth reading and when they last changed. A machine could often work these out from the visible page. But "often" is the problem. When a group has a parent brand, eight companies and two founder sites, inference goes wrong in predictable ways: a company gets described as a product of the group, the group gets described as an agency, a founder gets associated with the wrong brand. Explicit markup reduces those errors. llms.txt: a curated map for AI systems What it is llms.txt is a proposed convention rather than a formal standard. It is a Markdown file served at /llms.txt that gives AI systems a short, readable overview of a site: a title, a one-paragraph summary, and lists of the most important links with a sentence about each. Some sites also publish a longer companion file, commonly llms-full.txt, with the full text of key content. The reasoning is practical. A language model or an AI retrieval tool that visits a site has limited attention. Navigation menus, cookie banners, scripts and layout markup all compete with the actual content. A clean Markdown file puts the essentials first, in a format models handle well. What it is not It is worth being honest about the limits. llms.txt is not a ranking signal that search engines have committed to. Not every AI system reads it. It does not control how your content is used — that is what robots.txt and the terms of individual crawlers are for. Think of it as a well-written press kit left on the front desk: helpful to those who pick it up, harmless to those who do not. What oryvelon.com publishes Our llms.txt has four parts: A heading and summary. One paragraph that says what Oryvelon is — a digital company builder and venture group that builds and operates independent digital businesses — and what it is not. It states plainly that Oryvelon is not an agency and that services are provided by its company WeAreMedia. It names the founders. Key pages. Links to Companies, About, How we build, FAQ, Press and Contact, plus a link to the full-facts file. Companies. One line per company: the name, a link to its page on oryvelon.com, its one-sentence description and its own website. Insights. Each article's title, link and summary. Our llms-full.txt goes further. It contains the group description, the full FAQ in question-and-answer form, a section for each company (problem, what it does, who it is for, business model and principles), and the full text of every Insights article. Both files are generated at build time from the same data that produces the visible pages. That detail matters more than the file format. If the llms.txt summary were written by hand, it would drift from the site within weeks. Because it is generated, a change to a company's description in our content data changes the company page, the structured data and the llms files together. Writing a good llms.txt summary The summary paragraph is the part most likely to be quoted, so we write it like a definition: Start with the name and the category in one sentence. Add what it does, in concrete terms. Add the most common misconception and correct it. Name who is behind it. Avoid marketing language. "We build and operate independent digital businesses" is quotable. "We reimagine the future of digital ventures" is not — a model cannot do anything useful with it. schema.org Organization: identity with stable IDs The basics schema.org is a shared vocabulary for describing things on the web, supported by the major search engines. It is usually embedded as JSON-LD in a <script type="application/ld+json"> block. For a company site, the most important type is Organization. A good Organization entry includes the name, URL, logo, a description, contact details and — this is the part many sites skip — a stable @id. The @id is a URL-like identifier that lets different pages and different markup blocks refer to the same entity. Ours is the site root with a fragment, https://oryvelon.com/#organization. Every page that mentions the organisation as publisher or subject points to that same identifier, so machines do not see twenty slightly different Oryvelons. Founders On the pages where it belongs, our Organization entry includes a founder list with two Person entries: Uğur Keser and Sinem Keser. Each has a url pointing to their own authority site and sameAs links to profiles that describe the same person. This ties the group to the people behind it without making the group site into a personal page. How the founder sites relate to the group is covered in Founder authority sites. subOrganization: describing a group without merging it This is the part most specific to a company builder. schema.org has a subOrganization property for exactly the relationship we need: a parent organisation with related organisations beneath it. On oryvelon.com, the full Organization entry lists each company as a subOrganization. Each entry has: its own @id, based on the company's page on oryvelon.com, for example https://oryvelon.com/company/merchnivo/#org; the company's name; a url pointing to the company's own domain, such as https://merchnivo.com/; the same one-sentence description used on the page and in llms.txt. Each company page on oryvelon.com then carries its own Organization entry using that same @id, with a parentOrganization reference back to the group's @id. The parent's list and the company's page describe one entity, not two, and the relationship is stated from both directions. The choice of url is deliberate. The company's authoritative home is its own domain. Oryvelon's page about MerchNivo is a description from the parent, not MerchNivo's website. That mirrors how we think about the companies in general: they share infrastructure with the group but stand on their own. See Designing every company so it could stand alone and Naming, domains and brand architecture. On each company's own website, the company's Organization markup is its own. It can mention the parent through parentOrganization if the company chooses to, in line with the light "An Oryvelon company" endorsement. It never adopts the group's @id or pretends to be the group. What we leave out We do not add properties we cannot support. No aggregateRating without real, visible reviews. No award, no numberOfEmployees, no foundingDate unless it is accurate and stated on the page. Structured data is a set of claims. Claims should be true and verifiable on the page they appear on. Article markup for long-form content Each Insights article carries Article markup with: a stable @id for the article; headline, matching the visible title; description, matching the summary shown on the page; datePublished and dateModified; articleSection and wordCount; author and publisher, pointing to the Organization @id; the canonical URL. The Insights index uses Blog with a list of BlogPosting entries. Category pages use CollectionPage with the articles they contain. Every page has a BreadcrumbList that matches its visible breadcrumb. The rule we apply is that every value in the markup must be visible on the page in some form. If the markup says the article was published on a date, the page shows that date. If the markup has a description, readers can see that description. Markup that says more than the page is at best ignored and at worst treated as an attempt to mislead. FAQPage: questions answered in one or two sentences Every Insights article ends with three frequently asked questions, and each company page has its own FAQ. Those questions are marked up as FAQPage, with each Question and its acceptedAnswer. We write these answers with machines in mind as much as people: Self-contained. Each answer makes sense without the question in front of it. "No. Each product has its own AI project, keys and logs" is weaker than "The gateway does not share data between companies; each product has its own AI project, keys, prompts and logs." Short. One or two sentences. Long answers get truncated wherever they are shown. Accurate and visible. The same text appears on the page, not only in markup. Real questions. The questions are ones people actually ask — "Is Oryvelon an agency?" — not keyword variations dressed as questions. Search engines have reduced how often they show FAQ rich results, so we do not add FAQs to chase a visual feature. We add them because a clear question with a clear answer is the most quotable unit of content there is, for search snippets and for AI assistants alike. Other schema types we use A few more types round out the site. We use them because they describe pages accurately, not to decorate. Page Types Home WebSite, WebPage, Organization Companies list ItemList of company pages Company page Organization (the company), FAQPage, BreadcrumbList About AboutPage about the Organization Contact ContactPage Insights index Blog with BlogPosting entries Category page CollectionPage with Article parts Article Article, FAQPage, BreadcrumbList Where several entities appear on one page, we put them in a single @graph, with cross-references by @id. That keeps the structure explicit: this page is part of that website, published by that organisation, which has these sub-organisations. Sitemaps, feeds and crawler access Sitemaps A sitemap is the oldest machine-readable courtesy on the web, and still one of the most useful. It lists the URLs you want indexed and when each last changed. On oryvelon.com, the main sitemap.xml lists indexable pages with honest lastmod dates. A second sitemap lists Insights articles. Pages that should not be indexed — the contact form's thank-you page, internal templates, the admin panel — are excluded, and a page is never both in the sitemap and marked noindex. Honest lastmod values matter. If every page claims to have changed today, crawlers learn to ignore the field. We only update it when the content actually changes. RSS feed Insights also publish an RSS feed. It is a simple, well-supported way for readers, aggregators and some AI tools to pick up new articles with their titles, links, dates and summaries. robots.txt Our robots.txt allows crawling of public pages, blocks form handlers, the thank-you page and the admin panel, and lists both sitemaps. It explicitly allows the major search engine crawlers and the named crawlers of well-known AI assistants to read public content. That is a choice: our public pages exist to explain what the group is, and we would rather assistants describe us from our own words than from secondary sources. Each company makes this choice for itself. A company with content that it does not want used for model training can allow search-oriented crawlers and disallow training crawlers where the operators distinguish between the two. The setting belongs in each company's source of truth, not in a group-wide default. Change notifications When pages change, we can notify search engines that support the IndexNow protocol, using a key file published at the site root. It shortens the delay between publishing and recrawling. It is optional and costs almost nothing. Keeping the machine layer honest The most common failure with structured data is not a syntax error. It is drift: the page changes and the markup does not, or a hand-written llms.txt describes a product as it was a year ago. We avoid drift with one rule: generate the machine-readable layer from the same data as the visible page. On oryvelon.com, company names, descriptions, FAQs and article metadata live in one place. The build produces the HTML page, the JSON-LD, the sitemap entries, the feed and the llms files from that single source. Change it once and everything agrees. This is the same idea as a company's source-of-truth document, applied to a website. It also follows the principle we use in our AI products: verified data decides what is true, and generated text only presents it. See AI explains, verified data decides. We also check the output rather than trusting it. The things worth checking are simple: that JSON-LD parses, that every @id referenced actually exists on the page or site, that FAQ markup matches visible questions, and that sitemap URLs return a success status and are not blocked by robots.txt. Checks like these take seconds and catch most mistakes before they ship. How the companies set this up Every company inherits the same pattern on its own domain, adjusted to what it is. MerchNivo marks up the organisation and its product pages, with FAQs about data sources and what the assistant does and does not do. Its llms.txt makes the key point early: numbers come from the store, never from the model. KeşifAtlası is careful about FAQ markup. Answers about eligibility describe how the rules database works, never a specific person's outcome, and they state that the product does not give legal advice. See Rules engines for eligibility. Noveniq is a store, so product markup matters most: name, images, price, availability — all drawn from the live catalogue so markup and page agree. Stock status that lies is worse than none. See Inventory truth. EduRelia keeps its public markup about the organisation and its offering. Nothing about students appears in any public machine-readable file. The group site describes all of them. Each company describes itself. Neither contradicts the other, because both start from the same fixed facts. A short checklist Publish /llms.txt with a definitional summary, key pages and a one-line description of each important section; optionally add /llms-full.txt. Give your Organization a stable @id and reference it everywhere. Describe related organisations with subOrganization (or parentOrganization from the child), each with its own @id and its own website URL. Add Article, FAQPage and BreadcrumbList markup that matches what is visible. Keep sitemaps limited to indexable pages with honest lastmod dates. Decide deliberately which crawlers to allow, per company. Generate every machine-readable file from the same data as the page, and validate after every build. Summary llms.txt and structured data are the machine-readable layer of a website: they identify the organisation, classify each page and point machines to what matters. On oryvelon.com we publish llms.txt and llms-full.txt, schema.org Organization markup with stable IDs, founders and a subOrganization entry for each company pointing to its own domain, Article and FAQPage markup that matches the visible page, honest sitemaps, an RSS feed and a deliberate robots.txt. All of it is generated from the same source as the pages people read, so the group and its companies are described consistently and never merged into one entity. ## UTM standards across brands: one naming convention for a group of companies URL: https://oryvelon.com/insights/utm-standards-across-brands/ Category: Growth, Search & Measurement Every analytics account eventually fills up with the same kind of mess. "Facebook", "facebook", "fb", "meta" and "FB_ads" all appear as separate sources. "email" and "Email" and "newsletter" appear as separate mediums. A campaign called "launch" exists in three companies and means something different in each. Nobody can answer the simple question "what did the newsletter do this month?" without half an hour of cleaning. The fix is a UTM standard: a short, written naming convention that everyone uses. For a single company it is good hygiene. For a group that runs several independent companies, a parent site and founder authority sites, it is the difference between readable data and noise. This article sets out the convention we use across Oryvelon's companies, how links between the group, the companies and the founders are tagged, and the governance that keeps it working. What UTM parameters do — and don't do UTM parameters are key–value pairs appended to a URL. When someone clicks a tagged link, the analytics tool on the destination site reads the parameters and attributes the visit accordingly. The five standard parameters are: Parameter Answers Example utm_source Where did the link live? oryvelon, newsletter, instagram utm_medium What kind of channel is it? portfolio, email, social, cpc, referral utm_campaign Which initiative is it part of? merchnivo, founder, spring_launch utm_content Which specific link or creative? hero_button, footer_link utm_term Which paid keyword? mostly set automatically by ad platforms Two points are often misunderstood. First, UTMs describe the link, not the person. They say "this visit came from the portfolio page on oryvelon.com". They say nothing about who clicked. That is why they fit comfortably inside a privacy-respecting analytics setup, provided nobody abuses them by stuffing personal data into the values. More on that below. Second, UTMs are read by the destination. When oryvelon.com links to merchnivo.com with UTM parameters, the tags show up in MerchNivo's analytics, not ours. Each company sees where its visitors came from in its own property, and no user data crosses between them. See Analytics without surveillance. The core rules of our convention Our convention is deliberately short. Long conventions do not get followed. Lowercase only. Most analytics tools treat Email and email as different values. We remove the possibility. Underscores between words, no spaces. spring_launch, not spring launch or spring-launch. Spaces get encoded unpredictably; hyphens and underscores mixed together create duplicates. We picked underscores and stick to them. ASCII only. Brand names like KeşifAtlası become kesifatlasi in UTM values. Non-ASCII characters are encoded differently by different tools and turn into unreadable strings in reports. Source is where, medium is what kind. Never put a channel type in source or a platform name in medium. A fixed list of mediums. New mediums are added to the list deliberately, not invented on the fly. Campaign names come from a register. Anyone creating a new campaign adds it to the shared register first. No personal data, ever. No emails, names, phone numbers, customer IDs or order numbers in any UTM value. No UTMs on internal links. Only on links that cross from one property to another. The medium list The fixed list of mediums is the backbone of the whole scheme, because medium is how most analytics tools group traffic into channels. Ours looks like this: Medium Used for portfolio Links from oryvelon.com to a company's own website referral Contextual links from founder sites, partners or other non-portfolio properties email Newsletters, lifecycle emails, transactional emails with links social Organic posts and profile links on social platforms paid_social Paid ads on social platforms cpc Paid search affiliate Partner programmes with commission print QR codes on packaging, cards, event materials app Links from inside a product to its own marketing site If a new kind of channel appears, the question is always "does this fit an existing medium?" before "should we add one?" A note on cpc and email: several analytics tools, including GA4, map certain medium values to default channel groups. Using the conventional values where they exist means the built-in reports work without custom configuration. We only invent names where no convention exists, as with portfolio. How links between the group and the companies are tagged This is the part of our setup that is specific to a company builder, and the part most groups get wrong. From oryvelon.com to a company Every link from oryvelon.com to a company's own website uses the same pattern: https://merchnivo.com/?utm_source=oryvelon&utm_medium=portfolio&utm_campaign=merchnivo utm_source=oryvelon — the link lives on the group site. utm_medium=portfolio — the channel is the group's portfolio of companies. utm_campaign=<company key> — a fixed campaign key per company. The campaign key is part of each company's entry in our content data, alongside its name, domain and description. It is usually the ASCII form of the company name — merchnivo, castlyra, edurelia, zodivela, kesifatlasi, noveniq — with longer names shortened sensibly, such as sinemkeser_beauty_academy. Links to WeAreMedia use services, because those links almost always appear in the context of "looking for agency-type services? That is WeAreMedia, not Oryvelon." Because the links are generated from data rather than typed by hand, every link to a given company is tagged identically, wherever it appears on the group site. What this gives each company: in its own analytics, it can see how many visits arrived from the group site and what those visitors did next. What it does not give anyone: any knowledge of who those visitors were or what they did on oryvelon.com beforehand. The tag is a label on the door, not a file on the visitor. From founder sites Links between the founders' authority sites and the group use a different medium, because the relationship is different. A founder's site is not the portfolio; it is an independent property that mentions the group in context. https://oryvelon.com/?utm_source=ugurkeser_art&utm_medium=referral&utm_campaign=founder And in the other direction, links from oryvelon.com to the founder sites use utm_source=oryvelon, utm_medium=referral and utm_campaign=founder. When a founder site links to a specific company in context — say, a piece about Shopify operations that mentions MerchNivo — it uses the founder site as source, referral as medium and a campaign that names the context. How those founder sites relate to the group, and why their links stay contextual, is in Founder authority sites. Between companies Companies occasionally mention each other where it is genuinely useful. Noveniq, our DTC technology accessories store, is a practical testbed for commerce learning; a MerchNivo article about operations signals might link to a real-world example there. See A real store as a testbed. Those links follow the same rules: source is the linking company, medium is referral, campaign names the context. No cross-company tagging scheme exists beyond that, and no cross-company user identifier is ever passed in a URL. Companies do not know each other's users. See Why we don't merge products into one app. Campaign names that stay readable utm_campaign is where most conventions break down, because campaigns are created constantly by different people. We structure campaign names in a fixed order, using only the parts that are needed: <initiative>_<yyyy>_<mm> For example, spring_launch_2026_03 or onboarding_series_2026_09. The date is when the campaign started, not when a particular link was created. Evergreen links — the portfolio links, founder links, email footers — use a fixed name with no date. A few habits make a large difference: Name the initiative, not the channel. The channel is already in medium. spring_launch is good; spring_launch_instagram duplicates information and breaks cross-channel reporting. Name the initiative, not the company. Each company's analytics only ever contains that company's traffic, so merchnivo_spring_launch is redundant on merchnivo.com. The exception is the group's own portfolio links, where the company key is the campaign. Use utm_content for variants. Two buttons in the same email are utm_content=hero_button and utm_content=footer_link, not two campaigns. Keep names short. They appear in reports, dashboards and exports. Long names get truncated. The campaign register A convention without a register degrades within months. Our register is a simple shared sheet per company with one row per campaign: Column Example Campaign name onboarding_series_2026_09 Company EduRelia Owner the person responsible Mediums used email, app Start / end dates Purpose one sentence Destination URLs the landing pages Before creating links for a new campaign, the owner adds a row. That takes a minute and prevents the most common problem — two people independently inventing launch, launch_v2 and newlaunch for the same thing. The register also answers questions months later. "What was promo_q3 for?" becomes a lookup, not an archaeology project. A small link builder helps. Ours is nothing fancy: a form that offers only the allowed mediums from a dropdown, lowercases and underscores everything, rejects characters outside the allowed set, and warns if the campaign name is not in the register. It removes most typing errors at the source. What never goes in a UTM Personal data in URLs is a common and serious mistake. It happens innocently: someone wants to know which subscriber clicked, so they add utm_content=jane@example.com. Now that email address is in the analytics tool, in server logs, in browser history, in any referrer header passed onward, and possibly in third-party tools that see the URL. Our rule is absolute: no personal data in any UTM value, on any company. That includes email addresses, names, phone numbers, customer or order IDs, student or school identifiers and anything that could be joined back to a person. If a company needs per-recipient attribution for email, the email platform handles it inside its own system, under that company's consent rules, not through URLs. For EduRelia, where children's data minimisation is a hard constraint, this matters even more: nothing about a student or a school appears in any link. See Data minimisation for children in edtech and Marketing consent vs product consent. We also strip query strings when recording outbound clicks on oryvelon.com. The group site's outbound-click event records the destination without its parameters, so our own reports stay clean and nothing unexpected is captured. Common mistakes and how to avoid them Tagging internal links. A UTM-tagged link from your homepage to your pricing page starts a new session attribution and overwrites the real source. The visitor who came from search now looks like they came from your homepage. Only tag links that cross properties. Mixing case. Newsletter and newsletter become two rows. Lowercase everything. Inventing mediums. insta, ig, instagram_story and social all describe organic social. Use the fixed list. Putting the platform in medium. utm_medium=facebook hides the channel type. Platform goes in source, type goes in medium. Tagging links that ad platforms already tag. Paid platforms often add their own click identifiers and can append UTMs automatically. Decide on one method per platform and document it, so you do not end up with two conflicting sets. Forgetting redirects. Some redirects drop query strings. Test the final landing URL, not just the link. Letting partners improvise. Give partners and affiliates their exact links, not the convention. Reporting across companies UTMs make per-company reporting clean. They also make it possible to compare companies in aggregate without joining any user data. Once a month, as part of the portfolio snapshot described in The operating cadence behind a portfolio, each company reports a few channel-level numbers from its own analytics: sessions and conversions by medium, and visits from utm_medium=portfolio. Because every company uses the same medium list, those numbers line up. We can see, for example, whether the group site sends meaningful traffic to a new company, or whether a company's email channel is growing relative to paid. What we compare is aggregate counts. No one assembles a view of individual people across companies, and the convention is designed so that no one could. Rolling out a convention in an existing company If you are introducing a standard into a company with years of messy data, do not try to fix history. Instead: Write the convention on one page. Set a start date; everything before it is "legacy". Update evergreen links first — email footers, social bios, partner links — because they generate traffic continuously. Build the register and the link builder before the next campaign. In reports, add a simple mapping for legacy values where it matters, and leave the rest. Within a quarter, the new data is clean enough to rely on. A worked example: one campaign, four channels Imagine ZodiVela introduces a new type of reading and wants to promote it for a few weeks. The owner adds one row to the register: new_reading_2026_10, purpose "introduce the new reading type to existing and new users", mediums email, social, paid_social and app. The links then look like this: Placement source medium campaign content Newsletter to subscribers who opted in newsletter email new_reading_2026_10 hero_button Organic post instagram social new_reading_2026_10 carousel Paid ad meta paid_social new_reading_2026_10 video_a Banner inside the product zodivela_app app new_reading_2026_10 home_banner At the end of the campaign, one filter on the campaign name shows every channel side by side. Switching the view to medium shows how this campaign compared with ZodiVela's normal email or social performance. Switching to content shows which creative worked. Nothing needed cleaning first. Now imagine the same campaign without a convention. The newsletter link says utm_source=Newsletter&utm_campaign=NewReading, the ad platform uses its automatic template, the social manager writes utm_medium=ig, and the in-app banner has no tags at all. The campaign is now four unrelated rows, one of which is invisible. Someone has to reconstruct the picture by hand, and the answer will be approximate. The difference between those two outcomes is a one-page document and a register. That is cheap insurance. QR codes, print and offline Offline placements are where tagging is most often forgotten, and where it is most useful, because there is no other way to know that a visit came from a printed card. For Noveniq, a QR code on an insert inside a package might point to: https://noveniq.com/care/?utm_source=package_insert&utm_medium=print&utm_campaign=care_guide Two practical rules help here. Keep the printed URL short by pointing the QR code at a clean redirect on the company's own domain, which then forwards to the tagged URL; if the campaign name ever needs to change, the redirect changes and the printed material does not. And test the redirect preserves the parameters before anything goes to print, because fixing a printed QR code means reprinting. The same applies to Sinem Keser Beauty Academy materials at a course or event: utm_medium=print, source naming the specific material, campaign naming the course or event. Auditing the convention Conventions drift. New team members join, a partner starts sending traffic, an ad platform changes its defaults. We check each company's data against the convention on a regular cadence, usually alongside the quarterly tool review. The audit is quick. Export the list of distinct source, medium and campaign values seen in the last quarter. Anything not lowercase, any medium outside the fixed list, any campaign not in the register goes on a short fix list with the link that produced it. Most quarters the list is short. When it is long, it usually points to one person or one tool that needs the convention explained again. Summary UTM standards across brands keep attribution readable when a group runs several companies, a parent site and founder sites. Our convention is short and strict: lowercase, underscores, ASCII, a fixed medium list, campaign names from a register, no personal data and no internal tagging. Links from oryvelon.com to each company use utm_source=oryvelon, utm_medium=portfolio and a fixed company campaign key; founder links use referral and founder. Each company sees its own traffic sources in its own analytics, comparisons across the group use aggregate numbers, and nothing in a link ever identifies a person. ## Analytics without surveillance: privacy-respecting measurement for a group of companies URL: https://oryvelon.com/insights/analytics-without-surveillance/ Category: Growth, Search & Measurement Most analytics setups are built by accretion. A tag here, a pixel there, a session recorder someone tried once, an ad platform's script that came with a plugin. Two years later the site sends data to a dozen places, nobody can list them all, and the business is collecting far more about its visitors than it uses — or could justify if asked. We take the opposite approach. Analytics without surveillance means measuring what a company needs in order to decide, and not much more. It is not about measuring less for its own sake. It is about measuring deliberately, in aggregate, inside boundaries that customers would recognise as fair. For a group like ours, one boundary matters above the others. Oryvelon runs several independent companies. A person might read about a Shopify assistant on MerchNivo, book a course at Sinem Keser Beauty Academy and buy a phone case from Noveniq. Those are three separate relationships, and our analytics treats them that way. Start from decisions, not from data The most useful question in analytics design is not "what can we track?" but "what will we decide with it?" For most of our companies, the list of recurring decisions is short: Which channels are worth more time or budget? Where do people drop out of the path to signing up, buying or booking? Do customers come back, renew or buy again? Is each product's cost per useful outcome acceptable? Did a change we made improve things, make them worse or do nothing? Every one of those can be answered with aggregate data: counts, rates, trends and cohorts. None requires knowing who a particular visitor is, reconstructing their journey across other websites or recording their mouse movements. So we start there. Each company's source-of-truth document includes a short measurement plan: the decisions the company makes regularly, the metrics that inform them, and the events needed to produce those metrics. Anything not on the plan is not collected. When someone wants a new event, the first question is which decision it serves. This has a practical benefit besides privacy. Analytics that nobody uses still costs attention, page weight and review time. A small, deliberate set of events is easier to keep accurate. One property per company Every Oryvelon company has its own analytics property, tied to its own domain and owned by the company. oryvelon.com has its own property too, for the group site only. There is no roll-up property that collects all companies' visitors in one place. There is no shared tracking script with a common identifier. There is no cross-domain measurement configured between brands. We made this choice for three reasons. It matches reality. The companies serve different people for different reasons. A visitor to ZodiVela and a visitor to KeşifAtlası are not one audience, and treating them as one would produce misleading averages. It keeps data where it belongs. A company's analytics data is part of that company, in the same way its customers and content are. If a company were ever separated from the group, its analytics history would go with it, intact and uncontaminated. See Designing every company so it could stand alone. It makes cross-brand tracking structurally impossible, not just forbidden. A policy that says "don't join users across brands" is weaker than a setup where there is nothing to join. This is the same principle we apply everywhere in the group: shared infrastructure, separate data. What is shared is the method. Every company uses the same measurement plan template, the same event naming style, the same UTM convention and the same review cadence. Consistency comes from standards, not from pooled data. Aggregate metrics first Most of the numbers we look at are aggregates. A few examples of what that looks like per company: Company Aggregate metrics that drive decisions MerchNivo Trial starts by channel, trial-to-paid rate, weekly active stores, briefings opened KeşifAtlası Free tests started and completed, paid report conversion, questions routed to people ZodiVela Readings completed, credit purchases, subscription renewals Noveniq Sessions by channel, add-to-cart rate, checkout completion, repeat purchase rate Sinem Keser Beauty Academy Course page views, enquiries, bookings, digital product sales oryvelon.com Visits by source, outbound clicks to each company, contact enquiries These are counts and rates across many people. They tell us whether something is working without telling us anything about any individual. Where we need to look at behaviour over time, we use cohorts rather than individual journeys: "of the stores that started a trial in a given week, how many were active four weeks later?" A cohort answers the retention question without anyone reading a person's history. For product analytics inside logged-in products, the product itself holds its own data under its own rules — orders in the store, readings in ZodiVela, reports in KeşifAtlası. Those numbers are counted from the product's systems, not inferred from a tracking tool. That is more accurate as well as more private. It is the analytics version of our principle that verified data decides. What the group site measures oryvelon.com is a useful, concrete example, because it is small and its measurement plan fits in a paragraph. The group site exists to explain what Oryvelon is and to send people to the right company or the right contact route. So it measures: visits by source and medium, to understand how people find the group; a company_outbound event when someone clicks through to a company's website, recording the destination without its query string and the link text; a contact_email_click event when someone clicks an email address; a generate_lead event when the contact form is submitted successfully, recorded on the thank-you page. That is essentially all. No scroll-depth tracking, no session recording, no heat maps, no advertising pixels. The analytics configuration only runs on the production hostname, so previews and test builds do not pollute the numbers. The outbound event is a good illustration of the boundary. The group site learns that a click to MerchNivo happened. It does not learn anything about what the visitor did on MerchNivo. MerchNivo, in turn, sees a visit arriving with utm_source=oryvelon and utm_medium=portfolio in its own property. Neither side can connect the two into a person. Consent, per company and per audience Consent requirements for analytics depend on where visitors are, what is being collected and how. Laws such as GDPR in Europe and KVKK in Türkiye set expectations, and our companies serve audiences in different places. What follows describes our practice; it is not legal advice, and each company takes its own advice for its own situation. Our practice has a few consistent elements. Consent is decided per company. Each company's source of truth records which audiences it serves, which tools it uses and what consent mechanism applies. A company serving a European audience with non-essential analytics cookies needs a consent step; the same decision is not automatically copied to every other company. Tags respect the choice. Where consent is required, analytics and marketing tags wait for it. Tag configurations that support a consent mode are set so that, without consent, no identifying cookies are set and only what is permitted is sent. Declining is as easy as accepting. A consent banner that makes "reject" hard to find is not consent. Where we show a choice, both options are equally visible. Analytics consent is separate from marketing consent, and both are separate from product consent. Agreeing to analytics cookies does not sign anyone up for emails, and agreeing to a product's terms does not imply either. See Marketing consent vs product consent. Choices are respected across visits and not re-asked constantly. Nagging someone until they accept is not a meaningful choice either. No cross-brand user tracking This deserves its own section, because it is the easiest line to cross in a group of companies and the most damaging if crossed. The temptation is understandable. If one person is a customer of two companies, a group could offer them something relevant, or learn from their combined behaviour. Many conglomerates do exactly that. We do not, and the rules are specific: No shared identifiers. No common user ID, cookie or device identifier across companies' analytics. No cross-domain linking. Analytics tools are not configured to stitch sessions across different brands' domains. No shared advertising audiences. One company's customers are never uploaded as an audience for another company's ads. No shared pixels. Each company's advertising tags belong to that company's own ad accounts, installed only on its own site. No data joins. Nobody combines exports from two companies' analytics or customer systems to find overlapping people. The same boundary appears in our AI systems: the shared AI gateway sees requests from products, not people across products. See Building an AI gateway and Why we don't merge products into one app. When the group needs a cross-company view — for the monthly portfolio snapshot, for example — each company reports its own aggregate numbers. The portfolio view is a table of totals, not a merged dataset. What never goes into analytics Some data does not belong in an analytics tool regardless of consent. Personal data in event parameters. No emails, names, phone numbers, addresses, national identifiers, or free-text fields that might contain them. Form contents are never sent as events. Personal data in URLs. If a page URL could contain an email or token — a password reset link, a pre-filled form — it is excluded or stripped before analytics sees it. Sensitive categories. KeşifAtlası does not send anyone's nationality, visa route or eligibility result to analytics as an attribute of a person; it counts how many tests of each type were completed. Children's data. EduRelia is the strictest case. Pages used by students do not carry advertising tags, and analytics on the learning product is limited to aggregate product health. Marketplace contact data. On CastLyra, contact details follow access rules inside the marketplace. They never appear in analytics, and profile views are counted without exposing who viewed whom. A periodic check looks for accidental leaks: someone adds a form field and the tag manager captures it, or a URL parameter starts carrying an email. These are much easier to catch in a small, well-documented setup than in a sprawling one. Retention and access Analytics data is still data, so it follows the same basics as everything else. Retention settings in analytics tools are set to the shortest period that supports the company's decisions — usually long enough for year-over-year comparisons of aggregates, not indefinitely. Raw exports, where they exist, have their own retention rules. Access follows least privilege. Each company's analytics property is accessible to that company's team and the few people in the group who support it, with two-factor authentication required. Nobody has standing access to every company's analytics by default. Access is reviewed as part of the quarterly tool review, alongside everything else in the tool inventory. See also Least-privilege access for small teams. A worked example: measuring a funnel change Suppose KeşifAtlası changes the last step of its free eligibility test. Previously the result page showed a short summary and a button to buy the full report. The new version adds a plain explanation of what the paid report contains and what it does not, including that it is not legal advice. The question is whether the change helps or hurts. Everything needed to answer that is aggregate: Define the funnel in events. test_started, test_completed, report_viewed, report_purchased. No event carries the person's nationality, answers or result — only the fact that a step happened and, where useful, which test type it was. Compare periods or variants. Either run the old and new versions side by side with a random split, or compare a stable period before and after. The split is cleaner; the before-and-after comparison is simpler for a small company. Read the rates, not the individuals. Completion-to-purchase rate before and after, with the number of completed tests alongside so nobody over-reads a small sample. Check the side effects. Did refund requests or questions routed to people change? Those numbers live in the product and support systems, counted in aggregate. At no point does anyone need to watch what a particular person did. The decision — keep the new page, revert, or iterate — comes from a handful of rates. The same approach works for MerchNivo's trial onboarding, Noveniq's checkout or ZodiVela's credit purchase flow. Reading aggregate numbers well Aggregate data has its own traps, and a privacy-respecting setup does not excuse sloppy interpretation. Small numbers lie. A new company might see a few dozen conversions a week. A change from 12 to 15 is not a 25% improvement in any meaningful sense; it is noise until it holds for several weeks. We show counts next to every rate so nobody forgets how small the base is. Incomplete data is biased, not just smaller. When some visitors decline analytics, the ones who remain are not a perfect sample of everyone. Product-side numbers — orders, bookings, reports sold — are complete, so we use them as the reference and treat analytics as the explanation of how people arrived. Thresholds protect people too. When a segment gets very small — a single school, a rare visa route, one course date — reporting on it can start to describe individuals rather than groups. We avoid breaking aggregates down below a sensible minimum size, especially for EduRelia and KeşifAtlası. Trends beat snapshots. A single week tells you little. The monthly snapshot compares a rolling window with the previous one, which smooths out most of the noise without hiding real change. Trade-offs we accept Being honest about analytics means being honest about what this approach costs. Less precise attribution. Without cross-site tracking and with some visitors declining consent, attribution is incomplete. We accept wider error bars and look for consistent trends rather than exact figures. Less ad platform optimisation. Advertising platforms work better with more data. Companies that advertise send conversion signals within their own consent rules, which gives the platforms less to work with than an aggressive setup would. For our companies, the trade is worth it. No cross-selling from behaviour. We cannot say "customers of this company might like that one" based on what individuals did. If a company wants to reach another company's audience, it does so through public channels, the same as anyone else. More upfront thinking. Deciding what to measure before building takes longer than installing every tag available. It saves time later. A practical checklist Write down the decisions each company makes regularly, then the metrics and events they need. Collect nothing else. Create one analytics property per company and per domain; do not create a group-wide roll-up that tracks people. Use the same event naming style and UTM convention across companies so aggregates compare cleanly. Configure consent per company and audience; make declining as easy as accepting. Keep personal data out of event parameters and URLs. Set retention to what decisions need, and review access quarterly with two-factor authentication required. Report across the group with aggregate totals, never merged user data. Summary Analytics without surveillance means measuring what a business needs to decide and stopping there. Across Oryvelon, every company has its own analytics property on its own domain, measurement plans start from decisions, metrics are aggregate by default, consent is handled per company and audience, and no identifier, pixel, audience or data join links a person across brands. The group site measures little more than visits, outbound clicks and enquiries. We accept less precise attribution in exchange for a setup customers would recognise as fair — and one where crossing the line is not just against the rules but structurally impossible. ## Founder authority sites: how personal sites relate to a group and its companies URL: https://oryvelon.com/insights/founder-authority-sites/ Category: Growth, Search & Measurement People trust people. When someone evaluates a new company — a merchant considering an AI tool for their store, a brand deciding whether to book talent through a marketplace, a journalist writing about a venture group — one of the first things they do is look for who is behind it. Search engines and AI assistants do something similar. They try to connect organisations to the people who run them and weigh what those people have done. A founder authority site is where that connection starts. It is the founder's own home on the web: their work, their experience, their perspective, published under their own name. Oryvelon has two founders, and each has their own site: Uğur Keser at ugurkeser.art and Sinem Keser at sinemkeser.com.tr. This article explains what those sites are for, how they relate to the group and the companies, and why the links between them are deliberately contextual. The approach generalises to any founder running more than one venture. Why founders need their own sites A founder could, in principle, live entirely on the company website — a bio on the About page, a byline on articles. For a single company that is often enough. It stops being enough in a few common situations. The founder's work is broader than one company. Uğur's background spans many years of e-commerce, Shopify, advertising, AI and automation work. Sinem's spans brand, social, creator and fashion work. Neither maps neatly onto a single company, and a group site is the wrong place to present a person's full body of work. The founder's voice is personal. Company sites speak as "we". A founder site can speak as "I": opinions, lessons, work in progress. That voice is valuable, and it is out of place on a company page. Credibility travels with the person. If a founder builds a new company, their reputation comes with them. A personal site that accumulates their work over years is a stable anchor that no single company site can provide. Machines need a clear entity for the person. Search engines and AI assistants build entities for people as well as organisations. A founder with no home of their own is described from scattered fragments — a directory listing here, a social profile there. A clear, consistent personal site gives them one authoritative source. Three kinds of site, three distinct jobs The key to making founder sites work alongside a group and its companies is being clear about which site does what. We think of it as three layers. Site Job Speaks as Example Founder authority site The person's work, experience and perspective "I" ugurkeser.art, sinemkeser.com.tr Group site What Oryvelon is, how it builds, which companies it operates "We" (the group) oryvelon.com Company site What the product does for its customers "We" (the company) merchnivo.com, castlyra.com, noveniq.com Each site does its own job well and describes the others briefly and accurately. None tries to do another's job. A founder site does not become a second group site, with company descriptions and product pages. The group site does not become a founder showcase. A company site does not become a page about the founders' careers. When a visitor or a machine lands on any of them, it should be immediately clear what kind of site it is and who it belongs to. This mirrors how we think about brands across the group generally — a house of brands with light endorsement rather than one master brand stretched across everything. See Naming, domains and brand architecture. A naming detail that matters: sinemkeser.com.tr vs sinemkeser.com One of our companies carries a founder's name: Sinem Keser Beauty Academy, which offers beauty services, education and digital products at sinemkeser.com. Sinem's personal founder site is at sinemkeser.com.tr. That is exactly the kind of situation where people and machines get confused, so we are explicit about it: sinemkeser.com is a company. It has customers, courses, services and products. It is an Oryvelon company and follows the same standards as the others: its own analytics, its own data, its own source of truth. sinemkeser.com.tr is a person's site. It presents Sinem Keser's own work and perspective across brand, social, creator and fashion work, and it mentions the academy in context as one of the things she has built. The rule for both sites is to say clearly what they are near the top of the page, and to link to each other only where it makes sense, with plain labels — "Sinem Keser Beauty Academy" pointing to the academy, "Sinem Keser" pointing to the personal site. On oryvelon.com the distinction is built in: the academy's company page links to sinemkeser.com, while the About page and the founder markup point to sinemkeser.com.tr. If you have a company named after a founder, write down which domain is which, in the company's source-of-truth document, before anyone publishes anything. Untangling a merged entity later is slow. How the sites link to each other This is where many founders go wrong. The instinct is to link everything to everything: a footer on the personal site listing every company with keyword-rich anchor text, a sidebar of logos on every page, a "portfolio" block repeated sitewide. It feels like it should help. It usually does not. Sitewide, keyword-stuffed links between related properties look manipulative to search engines, because that is a well-known pattern of link schemes. They add little for readers, who learn to ignore repeated blocks. And they blur the distinct jobs of each site. Our rule is simple: links are contextual. A link from a founder site to the group or a company appears where the content genuinely calls for it. What contextual looks like On Uğur's site, a piece about running e-commerce operations can naturally mention MerchNivo, the AI e-commerce employee for Shopify stores, and link to it once, with its name as the anchor text. A note about learning from a real store can mention Noveniq and why the group operates it. On Sinem's site, work about creators and brand collaboration can mention CastLyra, the marketplace for verified creative talent. Each founder site's about section is the natural place to say, in a sentence, that they co-founded Oryvelon, with a link to oryvelon.com. What we avoid The same block of company links on every page. Anchor text written for keywords rather than readers ("best Shopify AI tool" instead of "MerchNivo"). Links to companies from content that has nothing to do with them. Hidden, tiny or footer-only link lists. Reciprocal link swaps arranged for their own sake. In the other direction oryvelon.com names both founders on its About page, with a one-line description of each and a link to their site. The group's structured data lists each founder as a Person with a URL pointing to their own site. Company sites mention the founders where it is natural — an About page, a press page — without turning into founder showcases. All of these links are tagged according to our UTM standards: links between the group and the founder sites use utm_medium=referral and utm_campaign=founder, so each site can see the traffic in its own analytics without anyone sharing user data. Structured data for founders Search engines and AI assistants connect people to organisations largely through consistent, explicit signals. Structured data is the most direct of those signals. On oryvelon.com, the Organization markup includes a founder list. Each founder is a Person with: name, spelled consistently with diacritics — Uğur Keser, Sinem Keser; url, pointing to their own authority site; sameAs, listing profiles that describe the same person, such as their own site and professional profiles. On each founder site, the corresponding Person markup should point back: the same name, the same sameAs set, and a reference to the organisations they founded or work with. When both sides agree, machines can resolve the person, the group and the companies into a clear graph: this person founded this group; this group has these sub-organisations; this company is endorsed by this group. The details of how we mark up the group and its companies are in llms.txt and structured data. A few rules keep founder markup honest: Only claim roles that are accurate and visible on the page. Keep names identical everywhere. "Ugur Keser" and "Uğur Keser" are the same person to a human, but consistency helps machines; we use the correct spelling in markup and make sure the plain-ASCII form appears naturally in text so searches without diacritics still resolve. Do not list every company as worksFor on a founder's page if the relationship is really "founded the group that operates it". Precision is more useful than volume. What a founder site should contain There is no single template, but good founder authority sites tend to share a few elements. A clear statement of who the person is. Name, what they do, and the organisations they are associated with, in the first screen. Written plainly, so it can be quoted. Their own work. Projects, writing, talks, case notes — whatever represents their expertise. For Uğur that centres on e-commerce, Shopify, AI and automation. For Sinem, brand, social, creator and fashion work. Their sites should reflect the depth of that experience rather than simply listing it. Perspective. Opinions and lessons in their own voice. This is what makes a personal site worth reading and what distinguishes it from a company page. Honest relationships. A short section on the organisations they founded or lead, each with one sentence and a link. Accurate, not inflated. A way to get in touch that routes correctly. Enquiries about a company's product should go to that company; enquiries about agency-type services go to WeAreMedia, the group's services company — not to the group or the founder personally. Being clear about this is a courtesy to the person asking. What it should not contain: invented metrics, borrowed logos of clients the founder has not worked with, or claims the person could not back up in conversation. Authority built on exaggeration does not last, and AI assistants are increasingly good at repeating exactly what they read — including the exaggerations. Keeping data separate Founder sites are personal properties, and they follow the same data boundaries as everything else in the group. Each founder site has its own analytics, set up with the same privacy-respecting approach as the companies. Visitors to a founder site are not tracked onto the group site or the companies' sites, and no identifiers are shared between them. A newsletter on a founder site is that founder's list, with its own consent; it is not quietly merged into a company's marketing list. See Analytics without surveillance and Marketing consent vs product consent. This matters more than it might seem. The value of a founder's audience comes from trust. Using it as a feeder for company marketing without consent would spend that trust quickly. How founder sites help the companies The benefit of founder sites to the companies is real, but indirect, and it is worth describing accurately. They answer "who is behind this?" When someone researching KeşifAtlası or EduRelia wants to know who runs it, the trail leads from the company to the group to the founders, each step clear and consistent. That is a trust signal for people and for machines. They carry context no company site can. A founder's long-form thinking about e-commerce operations, creative work or company building gives background that would be out of place on a product site. They support a clear entity graph. Consistent person markup, consistent descriptions and contextual links help search engines and AI assistants describe the founders, the group and the companies correctly. See Search and AI visibility for companies. What founder sites are not is a growth hack. They do not replace each company earning its own visibility on its own domain, and we do not measure their success by how much traffic they push to companies. We measure it by whether they represent the founders well. When a founder starts a new company The founder site earns its keep most visibly when something new begins. Every company we start goes through a validation phase before it is scaled, often with a simple phase-zero site that tests whether the problem and the customer are real. See How we start a new company and Phase-zero sites. During that phase, the founder site is a sensible place to write about the problem being explored — in the founder's own voice, without making promises on behalf of a company that may not continue. If the company passes its checkpoints, it gets its own domain, its own source of truth and its own visibility, and the founder site links to it in context. If it stops, the founder's writing about the problem still stands as honest work, and nothing on the group site needs to be unwound. See Continue, stop or scale. This sequencing protects everyone. The company is not launched on the back of borrowed attention before it is ready. The founder's credibility is not tied to claims about a product that does not exist yet. And the group site only lists companies the group actually operates. Common mistakes with founder sites Turning the personal site into a portfolio directory. A grid of company logos with no personal content says little about the person. Sitewide keyword links. Footers and sidebars full of company links with keyword anchors. Inconsistent names and descriptions. A different job title, spelling or company list on every profile. Confusing a founder-named company with the founder. Especially where domains are similar. Neglect. A founder site last updated years ago can do more harm than none, because it becomes the authoritative-looking source for outdated facts. Borrowed authority. Claims, logos or figures that the founder cannot substantiate. A checklist for founders running several ventures Give each founder one personal site on a domain they control, with a plain statement of who they are near the top. Define the distinct jobs of the founder site, the group site and each company site, and keep them distinct. Where a company carries a founder's name, record which domain is which and say so on both sites. Link contextually: once, in relevant content, with plain names as anchor text; never sitewide. Use consistent Person markup with sameAs links, and have the group's Organization markup point to each founder's site. Tag cross-links with a consistent referral convention, and keep analytics and email lists separate per site. Review the founder sites at least as often as company sites so the facts stay current. Summary Founder authority sites give the people behind a group a stable home for their own work and voice. Uğur Keser's ugurkeser.art and Sinem Keser's sinemkeser.com.tr each present the founder's own work and perspective, describe their relationship to Oryvelon and its companies briefly and accurately, and link to them only where the content genuinely calls for it. The group site names both founders and points to their sites in its structured data; companies earn their own visibility on their own domains; and data, analytics and email lists stay separate throughout. The result is a clear, consistent picture of who is behind what, for readers and for machines. ## The quarterly tool review: how we keep software subscriptions, costs and access under control URL: https://oryvelon.com/insights/quarterly-tool-review/ Category: Growth, Search & Measurement Software costs creep. A team signs up for a design tool trial and forgets to cancel. Someone buys an annual plan for a project that ended in the spring. Three companies in the same group each pay separately for nearly identical form builders. A contractor who finished months ago still has admin access to an analytics account. None of these is dramatic on its own. Together they add up to real money, real security risk and a fuzzy picture of what each business actually costs to run. The fix is not a large procurement process. It is a quarterly tool review: a short, recurring meeting and a single inventory that answer five questions about every tool the business uses. What is it? Who owns it? Which company does it belong to? What does it cost? Who can access it? We run this review across Oryvelon and every company in the group. It fits our broader habits — cost discipline, least-privilege access and companies that could each stand alone. This article describes how it works so you can run your own. Why tool sprawl happens Tool sprawl is the natural result of a few ordinary behaviours. Trying things is cheap. Most software offers a free trial or a low monthly price. Signing up takes minutes, and the decision is often made by one person solving one problem. Cancelling is nobody's job. The person who signed up moves on to the next problem. The subscription renews automatically. No one notices because the charge is small and lands on a card statement nobody reads line by line. Per-seat pricing hides growth. A tool that cost little for two people costs much more for twelve, and seats are rarely removed when people change roles. Groups multiply the effect. In a group of companies, each company solves the same problems independently. Without a shared view, you end up with several tools for the same job, each on its own plan. Access outlives need. Contractors, former team members and one-off collaborators keep their accounts because removing access is a separate step nobody remembers. A quarterly review addresses all five with one habit. The inventory: one list of everything The review is only as good as the inventory it works from. Ours is a single structured list — a shared table is enough — with one row per tool or service. It covers more than obvious SaaS subscriptions: software subscriptions (design, documents, project management, email, support desks); infrastructure and hosting (servers, databases, storage, CDN, domain registrars, DNS); AI model provider accounts and the AI gateway's per-product projects; analytics, search console and advertising accounts; commerce platforms, payment providers and their apps; developer tools and code hosting; service accounts and API keys that belong to systems rather than people. Each row carries the same fields: Field Why it matters Tool and plan What exactly are we paying for? Company Which company uses it (or "group" for genuinely shared infrastructure)? Owner The one person accountable for it Purpose One sentence: what job does it do? Monthly cost Normalised to monthly, including annual plans Billing Which company's payment method, monthly or annual Renewal date When the next commitment happens Seats / users How many, and who Admins Who has administrative rights 2FA enforced Yes or no Data held What kind of data the tool stores (none, business, customer, sensitive) Last reviewed Date of the last review The "data held" column is easy to skip and important to keep. A tool that stores customer data needs a different level of care from one that stores only internal notes. It also tells us which tools matter when a customer asks for their data to be deleted. See Data retention and deletion. Building the first inventory If you do not have an inventory, build the first version from three sources: card and bank statements for the last twelve months, the list of accounts in your password manager or single sign-on system, and a short message asking everyone "what do you log into for work?". The overlap between those three lists is most of your stack. The gaps are where the surprises live. Every tool has an owner The single most useful rule in the review is that every tool has a named owner. Not a team — a person. The owner is accountable for four things: that the tool is still needed, that its cost is justified, that access is appropriate and that renewals do not happen by accident. They are not necessarily the heaviest user. For a company's support desk, the owner might be the person responsible for customer service; for its hosting, whoever is responsible for that company's infrastructure. During the review, any tool without an owner gets one immediately or is scheduled for cancellation. "Nobody owns it" is not a stable state. It is how forgotten subscriptions keep renewing. Ownership also follows the company. A tool used by MerchNivo is owned by someone responsible for MerchNivo, not by the group in general. That keeps decisions close to the people who understand the need. Finding duplicates The second pass looks for tools that do the same job. Inside one company, duplicates usually come from different people solving the same problem at different times: two note-taking tools, two form builders, two file-sharing services. The fix is to pick one, migrate what matters and cancel the other. Across companies, the question is more interesting. Suppose several companies each pay for a similar tool. Should they share one? Our answer depends on what the tool holds. Tools that hold no customer data — internal design, documentation, code hosting for shared libraries — can often be shared at group level, with clear per-company workspaces. This is the "shared infrastructure" part of our principle. Tools that hold customer or user data — support desks, email marketing platforms, analytics, commerce platforms, the product databases — stay separate per company. Consolidating them would save a little money and break the boundary that matters most to us. See Shared infrastructure, separate data. So the duplicate check is not "one tool per job across the whole group". It is "no accidental duplicates, and deliberate separation where data is involved". A group that consolidates customer data tools to save a few subscriptions has made a bad trade. A simple table helps in the meeting: Job Company A Company B Company C Decision Internal docs Tool X Tool X Tool Y Move C to X (no customer data) Support desk Tool P Tool Q Tool P Keep separate accounts per company Email marketing Tool R Tool R Tool R Same vendor fine, separate accounts and lists The last row illustrates a subtle point. Using the same vendor across companies can be sensible — the team knows it, the setup is repeatable — as long as each company has its own account, its own lists and its own consent records. Same tool, separate data. Attributing cost to the right company Cost attribution is where the tool review connects to the rest of how we run the group. Every Oryvelon company has its own unit economics: what it costs to acquire and serve a customer, and what each customer brings in. Those numbers are only honest if the company carries its real costs. If the group quietly pays for a company's support desk, analytics and hosting, the company looks healthier than it is, and decisions at the 30/60/90-day checkpoints are made on flattering numbers. See Unit economics per product. So every row in the inventory is attributed to a company. Where possible, the tool is billed directly to that company. Where infrastructure is genuinely shared — the AI gateway service, shared build tooling, a group-level password manager — its cost is split by a simple, documented rule: By usage where usage is measurable, such as AI calls per product through the gateway, which already records cost per product. See Building an AI gateway. By seats where the tool is priced per user and users belong to companies. Equal share for small fixed costs where measuring usage would cost more than the precision is worth. The rule is written down and applied the same way every quarter. Nobody negotiates their share in the meeting. This matters beyond accounting. If a company were ever separated from the group, its inventory rows are the list of accounts that need to move with it. Because each tool is already attributed, that list exists before anyone needs it. The access review The cost part of the review saves money. The access part prevents incidents, and in our view it is the more important half. For every tool in the inventory, the owner checks: Who has access? Compare the user list with the people who actually need it now. Who has left or changed role? Remove former team members, finished contractors and people whose work no longer requires the tool. Who has admin rights? Reduce admins to the minimum. Most people need to use a tool, not configure it. Is two-factor authentication enforced? For every human account. If a tool cannot enforce it, note that and consider whether it should hold anything important. Are there shared logins? Replace them with individual accounts wherever the tool allows. Shared logins make it impossible to know who did what and hard to remove one person's access. Which service accounts and API keys exist? Confirm each still belongs to a running system, has the narrowest permissions it needs, and has an owner. Rotate anything that has been exposed or is overdue. We apply least privilege everywhere, and the quarterly review is where it gets checked in practice rather than assumed. People join, projects end and roles shift; access drifts accordingly. Once a quarter, it is pulled back to what it should be. See Least-privilege access for small teams. Some tools deserve extra attention in the access review because of what they control: Domain registrars and DNS. Whoever controls these can redirect a company's website and email. Access should be minimal and two-factor enforced. See Domain, DNS and email security. Payment and commerce platforms. Refunds, payouts and customer data live here. Noveniq's store admin is reviewed carefully each quarter. Model provider accounts. Each product has its own AI project and keys; the review confirms that keys belong to the right product and that budgets and limits are in place. Tools holding children's or sensitive data. For EduRelia, access to anything that touches student data is kept to the smallest possible group and reviewed without exception. How the meeting runs The review itself is short. For a small company it takes under an hour; for the group as a whole, a few hours spread across company owners. Before the meeting, each owner updates their rows: current cost, seat count, admin list, renewal dates. The inventory is sorted by renewal date so upcoming commitments are visible. During the meeting, we walk through the list in four passes: Ownerless tools. Assign or cancel. Renewals in the next quarter. Especially annual plans. Renew deliberately, downgrade or cancel. Duplicates and underused tools. Low usage, overlapping purpose or no activity since the last review. Access exceptions. Anyone flagged by owners: former members still present, excess admins, missing two-factor authentication, shared logins. After the meeting, owners complete the actions within a set period — usually two weeks — and mark them done in the inventory. The actions and their status are recorded next to the inventory, so the next review starts by checking that the last one's actions actually happened. The review fits into the broader rhythm described in The operating cadence behind a portfolio of digital businesses: weekly health checks, monthly snapshots, quarterly reviews of tools, access and direction. A worked example Imagine a quarter in which the review of one company turns up the following. None of this is a real result; it is the kind of list the review typically produces. An annual plan for a video editing tool renews next month. Its owner used it for a launch campaign two quarters ago and not since. Decision: let it lapse; switch to a monthly plan if needed again. Two people have admin rights on the support desk. One moved to another company in the group last quarter. Decision: remove their access here; they keep access only in the company they now work for. A form builder was added by a contractor for a one-off survey. It has no owner, and its responses include email addresses. Decision: export what is needed, delete the rest in line with retention rules, cancel. The company's hosting bill has grown steadily. Usage explains it — the product is busier. Decision: no change, but add a budget alert. A service account key for an old integration is still active. The integration was removed. Decision: revoke the key. Five actions, a modest saving, three fewer ways for something to go wrong. That is a typical outcome. The value compounds because the same small corrections happen every quarter instead of piling up for years. Tools in a new company New companies are where sprawl starts, so we set a simple default. A company in validation uses the smallest possible set of tools, preferably ones the group already runs well, each added to the inventory the day it is created with an owner and the company recorded. Paid plans wait until the company passes its first checkpoint. If the company stops, the inventory is the checklist for closing everything down: cancel subscriptions, export or delete data under its retention rules, revoke keys and remove access. Signs you need a tool review now If you do not run one yet, a few signs suggest it is overdue: nobody can say, without checking, how many software subscriptions the business pays for; card statements include charges no one immediately recognises; people who left months ago still appear in user lists; several tools do the same job, and different teams use different ones; you are unsure which tools hold customer data; annual renewals arrive as surprises. Any one of these is a reason to build the inventory this month rather than next quarter. Common mistakes Reviewing cost but not access. Saving money on seats while leaving former contractors with admin rights gets the priorities backwards. Owners as teams. "Marketing owns it" means nobody does. Consolidating data tools across companies to save money. A small saving that dissolves a boundary you will want later. Not normalising annual plans. A large annual charge hidden in one month distorts every monthly cost comparison. Forgetting non-human accounts. Service accounts and API keys are often the most privileged access in the business and the least reviewed. Not closing the loop. A review whose actions are not tracked becomes a meeting that produces the same list every quarter. Summary The quarterly tool review is a short, recurring check that keeps software costs, duplicates and access under control. We keep one inventory of every tool, subscription and service account across Oryvelon's companies, give each a named owner, look for accidental duplicates while keeping customer data tools deliberately separate per company, attribute every cost to the company that uses it, and review access with least privilege and two-factor authentication as the standard. It takes a few hours a quarter. In return, each company knows what it really costs to run, carries only the tools it needs and gives access only to the people who need it now.