GreenLight
Solves the clinician and Payer prior authorization faster, smarter AND safer.
Core trick
A reviewer agent cites tool-retrieved evidence per criterion, and a deterministic arbiter returns approve, deny, or insufficient, failing closed.
Description
Solves the clinician and Payer prior authorization faster, smarter AND safer.
README excerpt
# GreenLight **Live hackathon build — New Work.** All code in this repository is being written live during the hackathon. GreenLight is a clinical prior-authorization agent plus an eval harness: two pages sharing one engine. Per case, the engine parses a patient into a `PatientContext` (deterministic, no LLM), retrieves the clinician-selected CMS policy and decomposes it into criteria, has a reviewer agent judge each criterion with tool-grounded citations, optionally debates contested criteria, and hands the result to a deterministic arbiter that returns **APPROVE / DENY / INSUFFICIENT** — fail-closed, never a guess. ## Layout ``` adapters/ FHIR → PatientContext parsing (pure code) agent/ Claude Agent SDK engine: policy decomposition, reviewer, argument layer, arbiter api/ FastAPI backend — /case runs the engine and streams TraceEvents over SSE evals/ Eval harness against data/gold/ data/ patients (synthetic FHIR), policies, retrieval cache, gold labels ui/ Vite + React frontend (Motion animations) ``` ## Setup ```bash python3 -m venv .venv && source .venv/bin/activate pip install -r requirements.txt cd ui && npm install ``` All patient data is synthetic.