An EMR inbox management system that routes patient portal messages using documented clinical context — not just the message content. The same reported symptom can mean two different things depending on what a provider already documented and signed.
Built for the Abridge × Anthropic × Lightspeed Hackathon, July 18, 2026.
Most inbox triage reads only what the patient wrote. Intercept also reads what the provider already said — a documented threshold, a stated follow-up window, an explicit "call us if X happens" — and routes accordingly. The same message can produce different, correct outcomes depending on that context.
- Clinical / Refill — compares reported symptoms against documented thresholds
- Scheduling — resolves ambiguous requests ("I'm due soon," no date given) using real visit history; real Epic sandbox integration for actual slot availability and booking (human-confirmation gated)
- Billing — confirms what a charge concerns from context; never invents a dollar figure, structurally enforced
- Results — retrieves released values; never interprets what a value means, structurally enforced
- Fast safety-screening orchestrator (Haiku) runs on every message first
- Domain specialists (Sonnet) reason over message + real documented context
- Defense-in-depth: each specialist independently re-checks for missed emergencies
- Structural safety via Pydantic validators — certain dangerous outputs are impossible to represent, not just discouraged
- Multi-visit context selection for patients with more than one documented visit
- Real EHR integration: SMART on FHIR (standalone launch) + SMART Backend Services (JWT-based, no interactive login) against Epic's sandbox
Visit context is grounded in Abridge's provided synthetic-ambient-fhir-25 dataset (real transcripts/notes from synthetic patients) where available; scenarios are labeled with their actual data source, real or authored.
pip install -r requirements.txt
export ANTHROPIC_API_KEY=sk-ant-...
uvicorn main:app --reloadOpen http://127.0.0.1:8000.
- Evaluation labels are self-authored, not clinician-reviewed
- Repeated-contact thresholds are configurable starting points, not validated clinical constants
- No real pharmacy write-back exists; refill "actions" are audit-log entries, not real orders
- Scheduling scenario 6 area is fully synthetic — no real post-op scheduling data existed in the provided dataset
MIT