Discreet v0.1.0

First release of Discreet, a privacy gateway that lets teams use AI models on text containing Singapore personal data without sending that data to the model. Reviewed and approved by Michael Mendoza on 2026-10-06.

What it does

• OpenAI-compatible POST /v1/chat/completions: any OpenAI client works by changing base_url (checked with OpenAI's official Python library).
• Placeholders: NRIC/FIN (including mistyped), phone, email, card, postal code, unit number and date of birth become <NRIC_1>, <PHONE_1> … before the model sees them; values are restored in the answer. Cards are redacted and never restored. Values stay encrypted in memory for at most 10 minutes.
• Purpose required; automated eligibility and benefits decisions are refused (HTTP 403) before any model call.
• Tamper-evident audit log: every request, allowed or refused, with keyed fingerprints instead of text; discreet audit verify names the first changed record; CSV export.
• Money capped: the real model only for access-token holders, one pinned model, worst-case cost reserved against a daily budget (default US$0.20); everyone else gets a free mock.
• Demo page with four panes, and a container image ready for a mock-only public deployment.

Evidence

• 369/369 shared conformance cases (sg-pii-rules v0.2.0) in an independent Go implementation; 458 tests.
• Benchmark on 5,000 synthetic samples: 76.5% recall overall, 100% on every supported writing style, 92.3% precision; about 3 ms added per request, mostly the audit write (EVALS.md).
• Decision report: https://makoydev.github.io/discreet/m2-review/

Run it

docker build -t discreet . && docker run -p 8080:8080 -e DISCREET_HMAC_KEY="$(openssl rand -hex 32)" discreet

Then open http://localhost:8080/. The public deployment follows once the Fly.io account exists (docs/DEPLOY.md).