Sovereign AI
Self-hosted LLMs, RAG on your internal documents, local transcription: we deploy at your site the building blocks we run at ours.
Three situations
Contracts, client files, R&D, health data: your DPO or CISO is right to block it.
→ Sovereign document assistant: RAG on your documents, hosted in Europe or on your premises, no data leaves your infrastructure.
HDS hosting, local transcription of interviews (reports, consultations), AI embedded in your business tools.
→ We have already deployed self-hosted Mistral and Whisper on French health-data hosting certification (HDS) infrastructure.
Public sector, defence, OIV/NIS2: the sovereignty clause becomes a deal-breaker.
→ We design the architecture that meets it — and document it for your tender responses.
The proof
We don't resell American cloud with a European logo: our own infrastructure — 10 servers, our products, our AI agents — is self-hosted in Europe. We deploy for you what we use ourselves.
See the building blocks we track and deployThe approach
2 weeks. Requirements (GDPR, HDS, NIS2, tender clauses), open-source stack selection, sizing (CPU/GPU), TCO vs API, target architecture and deployment plan. Fully credited against deployment.
Sovereign document assistant in production (LLM + RAG + authentication), training included. Full platform or HDS: on quotation from the assessment.
EUR 1,990/month for HDS or dedicated GPU. Maintenance, model updates, monitoring, backups.
Entry tier
Open-source personal AI agents have become good enough that your teams install them on their own — usually on their laptop, usually wired into their work inbox. We deploy the same thing, but under control: on your infrastructure or a European server, with a deliberate set of permissions and a record of what the agent does.
An agent (OpenClaw or Hermes Agent, both MIT-licensed) on your infrastructure or a dedicated European server — levels N4 to N6 of our sovereignty scale.
→ Connected to your tools: mail, calendar, documents, through a dedicated service account.
Minimal permission scope, execution isolation, auditable action log, emergency stop, and how persistent memory is handled under GDPR.
→ The seven points of our hardening framework, closed and documented.
A written usage policy, traceable onboarding, a named contact. Article 4 of the EU AI Act already requires it — better to do it once, properly.
→ Extends into our AI literacy module.
One agent deployed and hardened on your infrastructure, connected to a first set of tools, with the usage policy and onboarding for the pilot team.
Maintenance, model and agent updates, quarterly review of permissions and logs.
This is the way in: if usage takes hold and the need grows into a company-wide document assistant, the pilot extends into a full sovereign deployment (from EUR 29,900) without throwing anything away.
In all honesty
If your data isn't sensitive and you already live in Microsoft 365, AI inside your tenant (EU Data Boundary) is often simpler and cheaper — that's our Excel Exit offer. The assessment tells you honestly which side you're on.
See the Excel Exit offerFAQ
European hosting (OVH), European open-weight models (Mistral), an audited open-source stack. There remain hardware dependencies (GPUs) common to the whole industry: we document precisely what is sovereign and what is not. Transparency is part of the offer.
For the vast majority of internal uses (document search, summarisation, transcription), yes. The assessment includes a test on YOUR documents before any commitment.
Beyond a certain volume — or as soon as a regulatory constraint exists — self-hosting wins. The assessment quantifies the tipping point for your case.
We do (monthly run), or your team after a skills handover — both are provided for.
The two approaches coexist, the assessment decides — see the “when not to choose self-hosting” block above.
That is the most common case, and it is exactly where we start. The pilot opens with a survey of what is already running — on which machines, with which access — then replaces the unmanaged install with a governed deployment. Banning it doesn't work: teams always find a way back.