ClearAuth

An autonomous agent pipeline that turns a doctor's clinical note into a filed, tracked prior authorization, with a human approval gate before anything reaches the insurer.

Overview

Prior authorization is the paperwork tax on American medicine: physicians spend around thirteen hours a week assembling payer-specific forms, hunting buried coverage criteria, and re-checking portals for decisions. ClearAuth automates everything on either side of the one moment that actually needs a doctor. A physician pastes a clinical note, and four autonomous agents extract the structured request, check it against the payer's coverage policy, assemble the prior-authorization packet, and audit it for compliance — then the pipeline deliberately stops. Nothing is filed until the doctor reviews the packet and approves; only then does a fifth agent drive a browser agent to submit the form on the payer portal and track the decision live. The model proposes, the clinician approves, the agents do the work in between. Built by our team of three for the Agents That Act hackathon.

What I built

• Five-agent pipeline with a human gate: extraction (note to patient, diagnosis, ICD-10, CPT, justification), criteria checking against payer policy, packet assembly, compliance audit with a pass/warn/fail verdict, and, only after physician approval, browser-agent submission to the payer portal with a captured confirmation ID.
• Fail-soft everything: every external integration is wrapped in a timeout with a deterministic fallback, so the full pipeline runs end to end with zero API keys; drop in a real key and that integration activates with no code change, and a bad key can never crash the run.
• Live agent timeline: every pipeline transition is persisted and broadcast over Server-Sent Events, so the clinician dashboard animates each agent step as it runs and updates the moment the payer's decision lands via webhook.
• Tamper-evident audit trail: each step appends an audit entry carrying a SHA-256 checksum over its own contents, re-derivable and verifiable after the fact, so the compliance trail can't be silently edited.
• Three-app monorepo: a Turborepo with the agent API, the clinician CRM dashboard, and a mock payer portal the browser agent actually fills and submits, all strict TypeScript with typed contracts shared across apps.

Tech

TypeScript · Next.js 15 · React 19 · LangChain · Claude · Turborepo · Server-Sent Events · Apify · Tigris · MCP · Supabase