Sovereign AI for industrial knowledge
Your senior service engineers are retiring. Their knowledge is in PDFs, tickets, emails and their heads.
Your technicians need answers in a machine hall with no signal. Your machine documentation cannot go into ChatGPT. And the people who know why a machine behaves the way it does are closer to retirement every year.
We compile what your company knows into a cited, auditable knowledge base that runs inside your own walls, and we build it with you rather than selling you a subscription.
The problem
The knowledge exists. It just isn't reachable.
Every machine builder we've worked with has the same four problems, and none of them are AI problems.
It's scattered
Manuals in one system, service bulletins in another, the real answers in tickets, emails, and one engineer's memory.
It's leaving
The people who commissioned the machines in 2009 are retiring, and what they know was never written down in a form anyone else can use.
It's needed offline
A field engineer in a basement plant does not have reliable connectivity, and the answer is needed now.
It doesn't need to go to a US-hosted model
Proprietary machine documentation, customer process data and tolerances are not going into a US-hosted model, whatever the terms of service say.
The approach
Open-weight AI, on your side of the firewall.
We compile your documents, logs and tickets into a cited, auditable knowledge base, and pair it with open-weight AI that runs on your own hardware: Apertus, the fully-open Swiss model, and Mistral. No US cloud, and no data leaving the perimeter. The wiki holds the facts with a source trail an auditor can follow; the model reads and reasons over them, on-premise or fully offline in the field.
How we work
Four steps. You own what comes out.
Find out what's answerable today
We start with a scoped engagement to find out where your machine knowledge actually lives, what condition it's in, and what a cited assistant could answer today. You get a written assessment and a working demo running on your own documents, not on a sample dataset.
Scope and duration are set per customer. It is deliberately small enough to approve without a committee.
scoped per customerAI compiles your knowledge into a cited wiki
We use AI to compile your documents, logs and tickets into a structured, cross-referenced wiki that regenerates as the sources change, with provenance behind every claim. The heavy generation runs in the cloud for speed, or fully on your premises when the corpus is too sensitive to leave the building.
We start from a core we've already built and harden it against your case. You are not paying us to invent a RAG pipeline from scratch, and you are not getting someone else's roadmap either.
cloud or on-prem generationLocal inference on hardware you own
You get a system, not a seat licence. Day-to-day queries run with local inference on the hardware you already have, so nothing leaves the building and there is no per-token cloud bill that grows every time your team uses it. It keeps working if we disappear, and gets better as it ingests more of your sources.
local inference, no token costIt gets sharper the more you use it
Every answer can be rated. A thumbs-down, or a correction from the person who actually knows, becomes a new cited entry in the corpus. The wiki fills its own gaps, so the system improves over time instead of going stale, and the knowledge compounds as your team works.
feedback becomes corpusThe blueprint
Four layers of automated operational memory.
Each layer builds on the one below it, turning what your systems already produce into a self-documenting system you can question and, if you like, talk to.
How a question becomes a cited answer
- 1 Logs
- 2 Parser
- 3 Ledger
- 4 Generator
- 5 Wiki
- 6 Mistral
- 7 Cited answer
Stack these layers and you have a self-documenting, queryable operational memory, built from what your systems already produce.
What we build on
The core already exists. The work is fitting it to you.
Sovereign Mind
Upload documents, get a readable cited wiki, and query it fully offline on a phone. Open-weight models (Apertus or Mistral), a citation on every answer, zero bytes leaving the device. This is the starting point for most engagements, tailored to the customer's corpus, vocabulary and workflow rather than shipped as-is.
The Automated Technical File
A pipeline that turns source code and operational logs into a cited, regenerable technical file, mapped to EU AI Act Articles 12, 13 and 14. Running in production on a physical machine with over 50,000 logged events.
Flotilla
Our open multi-agent orchestration framework, with cross-model peer review and circuit breakers. It's how a very small team delivers at this pace, and it's documented in a research paper on failure modes rather than in a brochure.
Why us
Thirty years of Swiss and German industrial software, and we publish what we find.
The founder architected machine control for semiconductor packaging at ESEC, integrated MES with Kistler's injection moulding sensors, ran R&D for thirty engineers at digitalSTROM, program-managed the digital platform at Komax, coordinated Komax machine data on Audi's production floor, and chaired the international working group that wrote the OPC UA Wire Harness Companion Specification through ARENA2036 in Stuttgart.
That is the difference between an AI vendor learning your industry and someone who has been in it since 1996.
We also publish our negatives
We tested our own fine-tuning approach against a proper control, found that retrieval did all the work and that the fine-tune actively hurt, and published the result with the full dataset. We would rather tell you what doesn't work before you pay for it.
Proof
Built and running, not slides.
Robot Ross · reference implementation
The Automated Technical File in production on a physical robot. 50,000+ logged events compiled into a cited, regenerable knowledge base with full provenance. The reference implementation.
Three published papers · all with released data
Failure modes in multi-agent fleets under subscription constraints. A four-arm ablation showing retrieval, not fine-tuning, carries the performance in a deployed RAG stack. A forward-only probe reading model disposition on smartphone-class open models, including the axis that looked real and wasn't.
Mistral Vibe Hackathon · Paris
Live log-triage pipeline built on Mistral models. Raw logs in, failure modes and fixes out.
Silicon Oracle · on-device, shipped
A full vision-and-language stack running entirely on an iPhone. Proof the on-device claim is real, wrapped in something that doesn't take itself too seriously.
Contact
Let's find out what your documents can already answer.
If you build machines and your service knowledge is scattered across systems and people, a short scoped audit will tell you more than a proposal will. We're a small team and we read every message.