OBHUT

Sovereign AI infrastructure

Your data never leaves your house.

Private LLM systems built on open-weight models that hold their own against the large providers. Run inside your infrastructure, under German and European law, certified rather than merely promised.

  • ISO 27001
  • ISO 42001
  • Europrivacy
  • BSI C5
  • BSI IT-Grundschutz
  • EU AI Act
Book a call

Usually answered the same working day.

Problem & answer

AI has long since arrived. Your legal team still says no.

Client files, patient records, personnel documents: exactly the material where AI saves the most time is the material nobody may send to a US provider. Professional secrecy under §203 StGB, processor obligations under Art. 28 GDPR, and from August 2026 the duties of the EU AI Act.

The result is rarely a clean no. It is half a yes: staff paste paragraphs into private chat windows because the official route takes too long. The data leaves anyway, just without a contract, without a log, and without anyone knowing.

The answer: the entire system stays with you. We build and run your private LLM system: the model, the RAG index, the MCP connections and, where it earns its keep, a model further trained on your own documents. Not a single token leaves your environment. Whatever the model knows about your organisation, it knows because you gave it.

Certification matched to the audience, not stacked for show:

Approach

What we actually do for you.

  1. 01

    Survey what exists

    We look at which documents you hold, where they sit and who is allowed to see them. Your existing permission model is the foundation, not an obstacle.

  2. 02

    Build the private system

    Open-weight model, a RAG index over your repositories, MCP connections into your line-of-business systems. All in your environment, all under your control.

  3. 03

    Train it on your house

    Where it pays off we train further: on your document structure, your clients, your workflows. Afterwards the model knows your file references, not just the language they are written in.

  4. 04

    Run it and prove it

    Monitoring, updates, model changes, and the evidence your auditors ask for. You get logs, not assurances.

Run wherever it suits you:

Same software, same evidence. Where it runs is your call, and it stays changeable.

What else we can do for you.

Document workflows without a line of code.

Bookable independently of the private AI system: a no-code surface where your own team assembles its workflows. A mail arrives, a document lands, a deadline approaches, and exactly what your people drew happens next.

The platform does not reinvent permissions. It inherits what your IT already maintains: accounts, groups, SharePoint shares. Anyone who could not see a document before still cannot see it inside a workflow.

Billing is per execution. No seat price, no minimum commitment. A workflow that does not run costs nothing.

Tap a node

Icons are data: arrow down means it comes in, arrow up means it goes out. Circles marked φ are functions.

Every node in the diagram is clickable. Tap or hover one for what it does. Drag to pan, pinch or use the buttons to zoom.

The circles marked φ are not program logic. They are instructions your own team writes in plain language: what to summarise, check or pull out. How that instruction gets carried out is the system’s job. Nobody on your side has to write code for it.

Impact

The time you get back.

Two sliders, two numbers from your own operation. The rest follows from them.

3.200hours per year
208.000 €saved per year
0,90 €cost per hour saved

Model: €0.12 per execution, €65 loaded cost per hour. Both replaced with your own figures in a written offer.

Frequently asked

What you are right to ask.

Where does our data actually end up?

Wherever you decide: your own data centre, a German provider, or your own cloud region. There is no back channel to us and no third-party model training. Where it runs is written into the contract, not into a brochure.

What happens to our documents during further training?

They stay in your environment. Training runs on your hardware or in your cloud instance, and the result is a model only you own. We take nothing away from it, including analytics.

Are open-weight models really good enough?

For summarising, retrieving, classifying and drafting against your own documents: yes, and the gap keeps narrowing. Where it is not enough we say so beforehand, not after the invoice.

How do you handle the certifications?

We name only what has been audited and produce the certificates on request. For work in progress we state the stage and the expected completion, marked as intent rather than as a seal.

What does it cost, and from when?

The AI system is quoted as build plus operation and depends on model size and where it runs. The workflow platform bills per execution. Both are bookable separately. Neither requires the other.

How long until the first workflow is live?

A first real workflow on real documents usually stands within a few weeks. The long part is never the technology, it is settling who may see what.