Last updated: 3 October 2026
This English translation is provided for convenience. In the event of any discrepancy, the French version prevails.
Hosting
The application and its database run in Paris (Fly.io, cdg region). Learners' data does not leave it for primary storage.
We do not describe this hosting as “sovereign”: Fly.io is an American company, and therefore subject to the CLOUD Act, even for servers located in France. The same applies to several technical components (authentication, emails), listed below. Sovereign hosting in the strict sense (a European provider with SecNumCloud qualification) is offered on a project basis, see three tiers of sovereignty.
AI models
The platform is not tied to any single provider: four are connected (Mistral, Anthropic, OpenAI, Google Gemini) and the model is chosen use by use, with no development required.
The three text-based AIs can be entrusted to Mistral, a French company:
- the persona, which plays the role opposite the learner;
- the assessment, which scores skills and writes the report;
- the orchestrator, which helps the trainer design their scenarios.
Voice mode can also be entrusted to Mistral: transcription of the learner's speech and the persona's voice go through Voxtral, Mistral's voice models. A complete session, text and voice alike, can therefore be processed end to end by a French AI provider. Other voice providers (Deepgram, Cartesia, Google, Inworld) remain available if your organisation prefers them.
Your own LLMs
Your organisation can route sessions through its own model: a Mistral or Azure account already under contract, or a model hosted on your premises (Llama, for example), exposed through a compatible API.
- You choose which uses are covered: the persona, the assessment, or both.
- Nothing leaves your endpoint if you so request. If your model is unavailable, the session stops rather than falling back to our providers.
- Your key is protected: stored encrypted (AES-256-GCM), never displayed again, usable only for your organisation.
- The network is controlled: HTTPS mandatory, address verified then pinned at connection time.
- Your costs remain your own: usage billed by your provider is accounted for separately.
Set-up is supported: we configure the connection with your IT department, then together we check the quality of the assessments produced by your model before opening it up to learners.
Data protection
- Isolation per organisation: every request is filtered on the user's organisation; a client never sees another client's data.
- Role-based access: organisation administrator, trainer, learner, each with a distinct scope.
- Automatic GDPR erasure: 30 days after an account is deleted, its sessions, transcripts, scores, reports and AI logs are permanently erased, by a process that runs every night.
- Backups: daily copy of the database, kept for 35 days, plus a monthly copy kept for 13 months. Restoration has been tested.
- Error monitoring hosted in the European Union (Sentry, Germany region).
Our role as a GDPR processor and our position on the EU AI Act are described on the Security, GDPR & EU AI Act page.
Three tiers of sovereignty
The first two tiers are available; the third is built on a project basis, with a dedicated timeline and budget.
| Tier | Text-based AI | Voice | Hosting and authentication | Availability |
|---|---|---|---|---|
| 1. European AI | Mistral | Voxtral (Mistral) | Paris, American providers | Available |
| 2. Your LLMs | Your endpoint, no fallback possible | Voxtral (Mistral) | Paris, American providers | Available, supported set-up |
| 3. Dedicated sovereign | Mistral or your endpoint | Voxtral (Mistral) | SecNumCloud provider (Scaleway, OVH, Outscale), European authentication | On a project basis |
Tier 3 also includes search within your course documents, which currently relies on an OpenAI model and would switch to Mistral.
Sub-processors
Standard configuration (tiers 1 and 2). The AI providers actually used depend on the choices made for your organisation.
| Function | Provider | Headquarters | Data location |
|---|---|---|---|
| Application and database hosting | Fly.io | United States | Paris |
| File storage and backups | Tigris | United States | Depending on configuration |
| Authentication | Clerk | United States | United States |
| Transactional emails | Postmark | United States | United States |
| Processing queue | Upstash (Redis) | United States | Paris |
| Error monitoring | Sentry | United States | Germany |
| Text-based AI | Mistral | France | Depending on configuration |
| Text-based AI (alternatives) | Anthropic, OpenAI, Google | United States | Depending on configuration |
| Speech transcription and synthesis | Mistral (Voxtral) | France | Depending on configuration |
| Speech transcription (alternative) | Deepgram | United States | Depending on configuration |
| Speech synthesis (alternatives) | Cartesia, Deepgram, Google, Inworld | United States | Depending on configuration |
Connect your own LLMs or go for the sovereign tier
Mistral or Azure endpoint, model hosted on your premises, requirements from your IT department: let's talk it through in concrete terms, with your real constraints on the table.
Talk to us