Projects

Things I've been building lately.

  1. Vigil AI

    Live Demo

    Clinical AI that knows when to act on its own and when to hand off to a human.

    • Built a multi-agent clinical pipeline (Monitor, Risk, Negotiation, Decision) in LangGraph, with confidence thresholds and a 4-zone autonomy framework that escalates to human review.
    • Ran structured evals across 50 test scenarios, debugging failures and iterating on prompts to bring the failure rate below 10%.

  2. Naya, Healthcare Voice AI Agent

    GitHub

    Voice scheduling for healthcare, where a missed escalation isn't an option.

    • Built a real-time voice agent (Pipecat/WebRTC) with identity-verified booking, RAG-grounded policy answers, and tone adaptation that shifts to warmer or faster replies based on caller affect and speaking rate.
    • Designed a deterministic guardrail layer that escalates emergency and crisis language directly in code, deliberately favoring false positives, since a wasted escalation costs staff seconds while a missed one could cost far more.
    • Wrote a 20-case behavioral eval suite that checks actual tool calls, not just replies, after catching the model claiming escalations it never made, plus 81 unit tests.