Axis 03 · Migration & modernization
Sovereign on-premise AI
In-house AI platforms (vLLM, Ollama) and search across your documents: your data never leaves your premises.
The challenge
Using online AI means sending your documents to foreign servers, outside your control.
NEXTBIT deploys open language models on your own infrastructure, served with vLLM or Ollama. We start from specific use cases, size the servers and GPUs required, then connect an assistant to your documents through semantic search with Qdrant. Every answer cites its sources and respects existing access rights: users only see the documents they are entitled to. Your data never leaves your premises.
What we do
- Use case study: document search, drafting, summarising, internal assistance
- Sizing of servers and GPUs according to usage
- Deployment of open language models with vLLM or Ollama
- Document assistant connected to your documents, with source citations
- Access control aligned with your existing rights, and request logging
Our approach
- 01
Assessment
Study of use cases, the documents involved, access rights and the available infrastructure.
- 02
Design and pilot
Model selection and sizing; pilot on a real use case, evaluated against criteria agreed with you.
- 03
Wave-based roll-out
Progressive opening to other departments and document collections, with access control at every step.
- 04
Operations and handover
Monitoring, model updates, documentation and training for your technical teams and users.
What you gain
- Your documents stay on your premises
- Answers with sources that users can verify
- No per-user subscription to a foreign service
- Open models, replaceable without lock-in
Frequently asked questions
Are open models good enough?
For document search, summarising and drafting, recent open models give good results, especially when they draw on your own documents. The pilot measures answer quality on your real cases before any wider investment.
What hardware do we need?
It depends on the models chosen, the number of users and the volume of documents. Sizing is part of the assessment: we specify exactly which servers and GPUs are needed, without reselling hardware, and you buy from the supplier of your choice.
How is personal data handled?
Everything stays on your infrastructure: documents, requests and answers. No data is sent to an external service. Access is controlled and logged, which makes it easier to comply with law 18-07 on personal data protection.