Sentinel - Public Service Voice AI Auditor

• 

  Contextual-integrity check as a Vadalog graph in Prometheux: the breach derived from explicit facts and a rule, with a visible chain.

• 

  Sentinel audits a Global Talent visa line: five cited breaches across EU and UK law, each traced to a live source.

• 

  Each finding cites a real legal source; the leaderboard ranks audited agents by breach count.

What it does

Sentinel autonomously audits AI voice agents used in public services for EU and UK compliance, and produces a cited, evidence-backed report; flagging the agents that never tell citizens they are talking to an AI.

Point it at a public-service voice agent and, with no further input, it grounds itself in live law via Tavily, judges the agent against six EU and UK compliance rules, and returns a verdict with every finding cited to a real legal source.

Inspiration

AI voice agents are being deployed across public services at speed: benefits lines, immigration queries, government helplines. Partnerships like ElevenLabs and DSIT are an early signal; many more public bodies will follow. The people these agents serve are often vulnerable: non-native speakers, those with low digital confidence, the elderly. When an agent does not disclose it is an AI, that is a breach of the EU AI Act's Article 50 and the UK's algorithmic transparency standards, and a real harm to citizens with no easy alternative. Right now, nobody systematically checks these agents. Sentinel is that missing inspection layer.

How I built it

• Tavily fetches the live text of the EU AI Act Article 50, the UK GOV.UK Service Standard and Algorithmic Transparency Recording Standard, and UK GDPR purpose-limitation guidance, so findings are grounded in current law, not hard-coded.
• Prometheux runs the contextual-integrity reasoning as a Vadalog relationship graph: was data collected for one purpose and reused for an incompatible one? It derives each breach from explicit facts and rules with a visible reasoning chain - auditable proof, not an LLM's guess.
• Claude (Anthropic) reads the call transcript and applies the disclosure and transparency rules, returning structured findings.
• ElevenLabs generates the demo target: a realistic Global Talent visa support line that never discloses it is an AI.
• Built in Cursor, on Next.js.

What I learned

The difference between a tool that thinks something is non-compliant and one that can show why. Relationship reasoning, not keyword matching, is what makes a compliance finding defensible to a regulator.

Challenges

Scoping a real, autonomous, multi-tool agent solo in a single day, and deciding what to cut to ship something genuinely working rather than half-built.

What's next

Live database leaderboard for monitoring agents over time, a programmatic provenance check, and auditing live phone lines directly.

Flags potential compliance gaps for human review. Not legal advice.

Built With

• anthropic-claude
• cursor
• elevenlabs
• nextjs
• prometheux
• tavily