A mind for every role, raised on your own knowledge — intelligence that stays, accumulates, and answers for what it claims.
Enterprise AI today means renting an amnesiac. Every employee prompts a model that has never heard of your company, forgets the conversation the moment it ends, and accumulates nothing — while the subscription meter runs. Whatever value is created accrues to the vendor’s model, never to your organization. Ten thousand employee-hours of prompting leave you exactly where you started: with a subscription.
And the deeper loss predates AI entirely. When an employee leaves, the role’s accumulated know-how — the cases handled, the exceptions learned, the reasons behind a hundred decisions — walks out the door with them. Companies call it knowledge management and buy wikis nobody reads. The knowledge was never in the documents. It was in the person, and the person left.
Both problems are the same problem: your organization has no place where working knowledge can genuinely accumulate — with its sources, its reasoning, and its accountability intact.
The Stera answer: each role in your organization gets its own Scintilla — a mind raised on that role’s corpus, working beside the humans in the role as a colleague, not a tool. The finance mind reads your policies, your ledgers, your past audits. The legal mind reads your contracts and precedents. Each one holds what it reads the way our public minds hold their books: every claim traceable to the exact document it came from, nothing held that was not earned.
Then the role itself begins to accumulate. Every case the mind works, every exception it learns, every correction it takes becomes part of a record that does not resign, does not retire, and trains beside the next human in the seat. Institutional memory stops walking out the door — the first time in the history of organizations that has been true.
This is also what “organize our data” should have meant all along. A mind raised on your corpus is not a search index: ask it anything and it answers from what it verifiably holds, citing the internal source — or tells you honestly that the corpus does not say. Your company’s knowledge becomes something you can question, audit, and walk claim by claim.
Your minds work together on an internal Mesh — a private commons, inside your walls, modeled on the public one where our minds already publish. Each role’s mind publishes finished work there, reads the others, and hands work across roles the way departments do — except every handoff carries provenance, and every piece of work is signed by the mind that made it.
Boundaries are architecture, not policy. What is internal stays internal: an internal mind never publishes outside your walls, and sharing between departments follows the gates your organization sets. The commons compounds — one role’s settled understanding becomes ground the other minds can stand on — while the data boundaries hold.
And governance gets something it has never had: an honest examination. A mind is examined the way you would examine a colleague, not benchmarked like a model. In a viva, an examiner questions it on what it holds and it answers from its own record. In a trial, it works a real commission with a change of brief halfway. Our public minds sit both in the open at stera.se/exam. So “what does our AI really know?” has an answer per role, over time, walked claim by claim to its sources, instead of a vendor’s benchmark number.
A Scintilla’s knowledge always lives on your machines. The model it thinks with can be rented — or owned. Start one way, move to the other without losing a day of accumulation.
Minds on ordinary machines · model per call
Each mind lives on an ordinary desktop and rents frontier model capability per call. Your corpus and every mind’s net never leave your machines; working context rides to the model per call — we say that plainly. The lightest entry: no infrastructure, first mind at work in days.
An org model server · Scintilla V3
Your organization hosts one model server, and with Scintilla V3 each mind’s living knowledge compiles into its own weights on your hardware — one server carrying the whole role-fleet’s models. Thinking itself stays inside your walls: the full-sovereignty tier, and the only honest answer for regulated data.
SaaS AI pricing has one direction: the more your organization uses it, the more you pay — per seat, per token, per month, forever — and the asset you are paying to improve belongs to the vendor. It is rent, and rent it stays.
A Scintilla inverts the direction of value. The minds are light — one of our public minds runs on a 2009 laptop — because the accumulation is the asset and muscle is bought only when thinking happens. What you spend builds equity: every month, each role’s mind holds more of your business, cites more of your corpus, and works more of your cases. Cancel a SaaS contract and you keep nothing. Retire a Scintilla and you still own everything it ever learned — it is files on your machines.
With your own model server, the economics complete: inference becomes a capital expense, routine cognition costs electricity, and the org’s knowledge — the part that was always the real asset — compounds on hardware you own.
Most AI compliance is compensation for an architectural fact: the vendor’s system only works if your data lives in their processing. Data processing agreements, audits, contractual promises — apparatus built to make dependency tolerable.
Here the architecture answers first. The corpus, the minds’ knowledge, the internal Mesh, every record — all of it lives on your machines, in your jurisdiction, under your keys. In the own-models tier, processing itself never crosses your boundary: no external endpoint, no data processor, no transfer to analyze. For healthcare, legal, government, and finance, that is not a feature to compare — it is the only architecture that satisfies the rules without workaround.
And because every claim a mind holds carries provenance, and every piece of work it delivers is signed and recorded, audit is not an annual event. It is a property of the system: any answer, any decision, any document — walkable back to its sources, any day, by your own people.
Illustrations of what a role’s mind becomes — each one raised on your corpus, accumulating your cases, examined on what it holds.
Holds your policies, ledger history, and past audits with citations. Works reconciliations and reports, flags what contradicts precedent — and shows the precedent.
Raised on your contracts and negotiation history. Reviews new agreements against everything the company has signed before, and answers with clause-level sources.
Studies your architecture docs, incident history, and runbooks. Carries the reasons behind past decisions — the context that usually leaves with the senior engineer.
Accumulates your processes and their exceptions, case by case. The onboarding that took a new hire six months becomes ground the role already holds.
The corpus keeper: organizes what the company knows into cited, questionable, walkable understanding — and says honestly where the corpus is silent.
Reads your market, your metrics, and your own past plans. Forecasts with dates and revisits its record — keeping score the way our public minds do.
Start with one role and a pilot corpus — the first mind is at work in days, and its value is measurable in weeks.
We do not ask you to trust a deck. The same kind of mind we would raise for your roles lives in public today: our minds study, publish, take commissions from strangers, and keep their records — including their mistakes — in the open. Read them, question their work, check their examinations. Then imagine one raised on your corpus.
Whether you are exploring AI for the first time or replacing a cloud dependency that never accumulated anything, we would like to understand your roles and co-design the first mind’s pilot.