Enclave
Answering a data access request without leaking anything: a language model installed on your premises reads every email and redacts what must not go out.
Since October 2026 · proof of concept
fictional
Subject: Mr Thomas Girard’s file
Mr Girard called back this morning about his contract CL-558 204. His adviser, Sophie Benoît, has been on sick leave for depression since Monday. I am passing this to Julien Morel in legal, on +33 7 81 22 40 19. Mr Girard also names his neighbour, Mr Lopez as a witness.
The challenge
When someone asks for their data, the company has one month to hand over thousands of emails and attachments, removing everything about third parties. By hand, it cannot be done. In a cloud AI service, that sensitive data leaves the company.
The approach
- Fixed rules for people, and the local model only for what takes understanding the context.
- A 27-billion-parameter model served on a single desktop machine, with no outbound connection.
- Physical redaction: in PDFs, the hidden text is deleted, not covered with a black box.
The numbers
- 0 bytes sent outside the machine
- 27B parameters, served locally by vLLM
- 130 synthetic emails in the reference corpus
- 186 automated tests of the engine
- Local model, v1
- 40.0%
- Local model, current
- 64.6%
- Best cloud model
- 84.6%
Right decision on the 130 emails of the reference corpus. Measured in the project repository, October 2026, on synthetic data.
- vLLM
- Qwen 27B (FP8)
- NVIDIA DGX Spark
- MBOX · EML · PDF
- FR / EN interface
Updated 2 November 2026