Built solo at The Future of Agentic AI in Healthcare hackathon (Abridge x Anthropic x Lightspeed), July 18, 2026.
An agentic judgment layer that turns inpatient clinical data into proactive family updates — and knows when to say nothing. Green-tier facts become plain-language texts to the care circle. Actionable items promise a human call. Acute events lock the channel: bad news only ever travels by human voice, and the agent's job is to make sure the nurse walks into that call prepared.
synthetic-ambient-fhir-25.json— synthetic dataset provided by Abridge for hackathon use. Not my work.- Everything else in this repo — the judgment/verification pipeline, tier logic, channel lock, and both interfaces — was designed and built during the event, today. See commit history.
- Demo presents two encounter days from the synthetic record with illustrative day labels, ordered stable-then-deterioration to reflect a typical clinical course; underlying data dates are in the code.
- FHIR data loader over the Abridge synthetic record (patient 7bd9e5b0, COVID inpatient admission)
- Two-pass AI judgment layer: tier classification (green/yellow/red) with per-fact reasons, asymmetric uncertainty rule (uncertain → more restrictive tier)
- Tier-based oversight: green auto-sends, yellow requires nurse approval of a fixed callback template, red locks the family channel
- Per-recipient generation: plain-language family updates, Spanish-first with English toggle
- Reply agent with three-way resolution: released-ledger answers / occurrence-only event answers (with honest absence) / hard-coded uniform routing (byte-identical regardless of question severity)
- Red-day agentic call coordination: holding message + time selection, feeding the nurse escalation card ("what Teresa already knows")
- Nurse view, iPhone-framed family view, live status pipeline
python3 -m http.server 8080 in the repo folder → localhost:8080 → paste an Anthropic API key (held in memory only) → click the days.
Single patient, two demo days (labels illustrative, ordered stable→deterioration; data dates in code), dictation seeded (marked "simulated" — production input is nurse voice / ambient capture), SMS simulated in-browser.