Sovereign AI for regulated institutions
We compile your institution's knowledge into a cited, auditable wiki, bake its judgment into a sovereign model you own, and run the whole machine inside your own walls.
Specialized AI for banking, healthcare, and engineering — where the answer has to be right, accountable, and yours. Facts stay retrievable. Judgment gets baked in. Nothing leaves the perimeter.
The idea
Your documents are the source code. The wiki is the binary.
Most systems force the model to be the database — re-reading the same documents on every query, improvising what it knows. We split the work the way it should be split: facts live in a compiled, cross-referenced corpus the model consults, with a source trail an auditor can follow. Judgment lives in the weights. Facts change, so they stay retrievable and citable. Judgment is stable, so it gets baked in.
The machine
One system. Three things it does — each already running.
We sell the machine, not the model. Not a chatbot, and not a service you pipe your data into — a pipeline that turns one institution's knowledge into AI it owns and operates.
Knowledge into a cited wiki
We ingest your documents, code, and operational logs and compile them into a structured, cross-referenced knowledge base that regenerates as sources change — with provenance behind every claim.
facts stay retrievableJudgment into a model you own
We fine-tune your institution's reasoning into an open model — built on Apertus, the fully-open Swiss base, or Mistral's open weights — so it brings expert judgment without holding facts it shouldn't.
judgment, not factsInside your own perimeter
The model and the wiki run on your infrastructure — on-premise or on-device. No data crosses the boundary, no dependency on an outside API staying available or affordable.
nothing leavesWhat it changes
Put knowledge in the right place, and the everyday math changes.
Your context window stays free
The model already carries your judgment and the wiki holds your facts, so you stop re-loading huge instruction files into every prompt. The room goes to the actual task.
Guidelines distribute themselves
Company policy lives in the weights and the wiki. Every model and every agent applies the same rules — no chasing down who has the latest handbook.
Less to load, less to spend
No megabyte instruction dumps on every call means fewer tokens, lower cost, and faster answers — and large instruction files stop quietly degrading as they grow.
Knowledge that compounds
The wiki sharpens as it ingests new sources, and the model can be re-baked when it's worth it. The machine gets better the longer it runs — and it stays yours.
No cold start
A fresh model instance is born knowing the institution. No re-learning the same context at the top of every session — the knowledge is already there.
Cost you can predict
You run an owned system on fixed infrastructure, not a meter that climbs every time more of your staff use it. No per-token billing surprises.
Why it's different
Built for institutions that can't send their data anywhere.
Sovereign by default
The model and the knowledge run inside your walls. The right answer for anyone who legally or commercially cannot ship data to a third-party cloud.
Auditable, not a black box
Because facts stay in a cited wiki rather than dissolved into the weights, every answer traces back to a source a regulator can inspect.
You own the machine
We hand over a system, not a subscription. It compounds in value the longer it runs — and it stays yours.
No vendor to depend on
It keeps working regardless of an outside provider's pricing, terms, or sudden access limits — including export-control directives that can strand a closed model overnight. It can run fully air-gapped.
Your IP stays in-house
Your data, code, and know-how are compiled and baked on your own hardware. Nothing is sent to an outside model, and nothing trains someone else's.
Harder to hijack
Judgment held in the weights is harder to override than instructions pasted into a prompt — a smaller surface for prompt injection. Harder, not immune; we're honest about that.
Proof, not slides
Each part is built and running today.
The thesis isn't a pitch deck — it's a set of systems in production, plus products and demos you can put in your hand.
Robot Ross · the wiki, in production
Our Automated Technical File compiles 50,000+ operational events into a cited, regenerable knowledge base — live, with full provenance. The reference implementation of the whole thesis, running on a physical machine.
Flotilla · multi-agent engineering
Our open orchestration framework coordinates a fleet of models with cross-checked peer review — documented in a research paper on failure modes in subscription-constrained fleets.
Read the workSovereign Mind · the product
Upload your documents, get a readable wiki, download the knowledge base to your iPhone, and query it fully offline — open weights, citations on every answer, zero bytes leaving the device. The machine, productized.
Request early accessSilicon Oracle · on-device, shipped
A full vision-and-language stack running entirely on local hardware — four oracle voices, no cloud, no account. Proof that a serious on-device model fits in your pocket, wrapped in something that doesn't take itself too seriously.
Try itFrom the field
Mistral Vibe Hackathon · Paris
We brought Robot Ross and built a live log-triage pipeline on Mistral models — raw logs in, failure modes and fixes out, over an API. Stress-tested the on-device stack in front of people who break demos for fun.
Next
Sovereign, on-device AI — Zurich talk
A walk-through of what it actually takes to run a full RAG stack offline on a phone, on open Swiss and European weights. Details & date →
Where it fits
Domains where the answer has to be right — and accountable.
Get in touch
Let's talk about your institution.
Whether you're evaluating sovereign AI, building a pilot, or just curious — we're a small team and we read every message.