Local AI
Language models in your own house.
Open models are good enough for real work — chat, searching documents, code. The difference from cloud AI: your data stays where it belongs. With you.
Local AI means running open language models like Llama, Qwen or Kimi with tools like Ollama, Open WebUI and RAGFlow on your own hardware or EU servers — prompts and documents never leave the company.
What I build
- Model and hardware advice: what your use case actually needs — from a workstation to a GPU server
- Chat and assistance with Ollama and Open WebUI, usable across the team
- RAG over your own documents with RAGFlow: answers from your knowledge, not the internet
- Coding assistance with local models (Llama, Qwen, Kimi & co.)
- GDPR embedding: local AI does not replace the process — it makes it easier
- Operations, updates and model maintenance afterwards
Cloud AI or AI on your own premises?
| Cloud AI (US provider) | Local AI | |
|---|---|---|
| Data flow | Inputs leave the building | Everything stays on your infrastructure |
| GDPR | Processing agreement + third-country transfer to sort out | No third-country transfer — a much simpler assessment |
| Costs | Subscription or per-request usage, forever | Hardware once, plus operations |
| Capability | Frontier models, always current | More than enough for many everyday tasks |
| Confidentiality | Contract and trust | Technically guaranteed — the data doesn’t leave |
How it works
First call
Free: which use case, which data, which expectations?
Pilot
One use case, real data, a clear success criterion — before you invest.
Rollout
From pilot to team tool: permissions, interface, integrations.
Operations
Updates, model changes, monitoring — AI is software, it wants maintenance.
Who this is for
For companies that want to use AI without sending customer data to US providers — and for everyone who first wants a manageable pilot to see what local models can really do.
Frequently asked questions
Is local AI good enough? +
For many tasks, yes: text, summaries, search over your own documents, coding assistance. Whether it is enough for your case, a pilot with real data shows faster than any slide deck.
What hardware do we need? +
That depends on the model and the number of users — from a single workstation to a dedicated GPU server. Sizing is part of the advice; you buy once the need is clear.
Are we automatically GDPR-compliant with local AI? +
No — local AI solves the transfer problem but does not replace the process: purposes, legal bases and deletion concepts are still required. The advantage: with data in-house, that process gets much simpler.
What does getting started cost? +
The pilot is deliberately cut small and gets a clear effort estimate up front — after the free first call you know what to expect.
From the blog
Setting Up Matomo Cookieless and GDPR-Compliant
Matomo runs without a cookie banner when configured right: cookies off, IPs shortened, retention set. The exact settings — and their honest trade-offs.
High-Risk AI Under Annex III: Are You Affected?
CV screening, credit scoring, exam grading: Annex III of the AI Act turns everyday AI into high-risk AI. The eight areas, and what deployers actually owe.
AI Literacy Duty: What Article 4 Actually Requires
No mandatory course, no certificate: the AI Act's literacy duty is smaller than its reputation. What belongs in it — with an outline for the internal training.
AI yes — but without the data leak?
In a free first call we find the use case worth piloting.