Sovereignty · Anonymisation filter
Your sensitive data never leaves in the clear
When an agent has to call an external AI model, a pre-processing filter first pseudonymises the direct identifiers — then the answer is re-associated locally with your context. Try it below: the demonstration runs entirely in your browser.
A local demonstration: your text stays in this browser — nothing is sent to a server.
Sent to the external model (anonymised)
Correspondence table
| Token | Type | Original value |
|---|
Detection of direct identifiers (email, phone, IBAN, social security number, company registration, card, IP) by validated patterns; names and addresses on a best-effort basis. A filter reduces the risk, it does not remove it: for the most sensitive data, the agent relies on local inference (no external call).
How it works
- 1. Detection & pseudonymisation — before any external call, direct identifiers are replaced by stable tokens ([EMAIL_1], [IBAN_1]…).
- 2. Processing — only the pseudonymised text is sent to the model; the correspondence table stays local and is never transmitted.
- 3. Re-association — the answer is re-associated locally with your real values. In strict mode, the masking is permanent (irreversible).
- The most sensitive data — the agent switches to 100% local inference: no data leaves your environment.
A use case in a regulated sector?
Health, finance, legal: let's scope the right level of sovereignty together.