Your own AI. Your house. Your data.
A fully local language model on your own hardware — open source, offline-capable, with its own knowledge base. Capture the value of AI without handing trade secrets and personal data to third-party clouds. The model is downloaded once; after that the system can run entirely offline.
Four building blocks — all running locally.
All open source — transparent, auditable, replaceable. No black-box vendor.
Mistral
The open-source language model (Apache 2.0) from EU provider Mistral AI — multilingual, generates the actual answers, runs fully locally.
Ollama
The runtime that loads and runs the model — uses the built-in Apple GPU on the Mac. One command loads or swaps models.
Open WebUI
A browser interface like ChatGPT — user accounts and permissions, multilingual, with the built-in knowledge base for your own documents.
RAG + Embeddings
Your own documents and emails are turned into searchable vectors locally (nomic-embed-text) and given to the model as context — no upload, no retraining.
Data sovereignty, predictable cost, independence.
Data stays in the house
No transfer to third parties. Trade secrets, personal and client data stay under your own control.
No per-request fees
No token price, no per-seat licence. Runs on existing hardware — predictable fixed cost.
No vendor lock-in
Open source, interchangeable models. No dependency on one provider, its prices or terms.
Knows your knowledge
RAG binds in your own documents and emails — answers with your context, not generic general knowledge.
Multilingual
Ask and answer in any language — practical for international teams and correspondence.
DSGVO by Design
Processing in-house / in the EU, no third-country transfers, no telemetry — data protection is built in.
Connect any of your own sources — everything stays in the house.
Help desk with instant answers
Customers describe their issue by voice; the model delivers a solution from manuals, FAQs and past tickets — 24/7, multilingual.
Assistant for technicians
On-site queries for installation, wiring and repair instructions — from technical docs, schematics and manuals. Also for training and onboarding.
Offer & document assistant
Draft offers and reports from templates and history, check contracts and tenders against requirements, summarise long documents in seconds.
Secure — and provable.
Not “we promise” but architecturally enforced and verifiable.
Configuration
Everything that phones home is switched off — no telemetry, no external AI services, no web calls. Local accounts and access rights.
Architecture
Airgap possible: the system can run fully disconnected from the internet — data leakage is then physically impossible.
Proof
Verifiable, not a matter of faith: a network capture shows zero outbound connections. All on your own, encryptable hardware.
Not either/or — the right tool for the data.
An internal LLM is not a replacement for everything. Transparency about the trade-offs is part of the deal.
- Data protection: everything stays in the house / EU.
- Fixed cost, no per-request fee.
- Fully offline / airgap-capable.
- Strong for everyday work and knowledge search on your own data.
- Top-tier reasoning: large cloud models are stronger on highly complex tasks.
- No live web — the knowledge has a cut-off date.
- Like any LLM it can hallucinate — RAG and a checking eye stay important.
- Operation, updates and concurrent-user capacity are your own responsibility (hardware-bound).