An AI assistant that answers "does the EU AI Act apply to me?" with citations, runs a bias examination and keeps the evidence, and gives every answer a mark anyone can verify.
1. Answers cited from the law. Hybrid RAG over the consolidated EU AI Act, Digital Omnibus included. Every claim cites the article it comes from, and answers are checked by an evaluation judge from a different model family (faithfulness 0.95, mean of the latest runs).
2. From obligation to evidence. For the bias-examination duty in Art. 10(2)(f), a tool-calling agent runs a Fairlearn check on a benchmark dataset (demographic parity and equalized odds, with a pass or flag verdict). It stores the result as dated evidence against the control, and compiles a compliance report that traces every obligation to its control, test and evidence.
3. A law that watches itself. RegWatch, a LangGraph agent, monitors official EU sources and investigates changes with a second model family. It proposes registry updates that only a human can approve.
4. Transparency applied to itself. Every answer and export carries a signed, machine-readable AI-generated mark (Art. 50(2)). Anyone can verify it on a public page with no account.
5. Run like a product. Deployed on Fly.io and Vercel, with magic-link sign-in behind an allowlist, per-person spend caps, layered safety guardrails, CI and production probes, about 1,200 tests, and written-up incidents.
What's next: testing a company's own hiring outcomes from aggregate counts only, using the impact-ratio method NYC Local Law 144 prescribes (the engine already reproduces a published LL144 bias audit to within 0.002), and Flow, which brings the same checks into engineers' tickets.
Try it and share feedback
- Try it now, no account needed: open https://fairaudit-six.vercel.app/transparency, click "Try it with a live FairAudit report", verify it, then change one character and watch the verification fail.
- Use the full assistant: https://fairaudit-six.vercel.app is a private beta. Email stela@fairaudit.ai with a line about who you are, and I'll send you an access code.
- Feedback welcome: a question it got wrong, a law it should know, or something confusing. Email stela@fairaudit.ai. Every piece of feedback is logged and tested, and it gets a decision.
