An enterprise-grade AI orchestration system designed to automate the medical chart review process. AutoPA focuses on deterministic execution, safety, auditable reasoning, and Human-in-the-Loop (HITL) handoffs to solve the $35 Billion prior authorization bottleneck in modern healthcare.
AutoPA is a full-stack State-Machine Agent application that acts as a digital clinical investigator. Instead of a human nurse manually digging through hundreds of pages of unstructured medical records, this agent reads complex insurance policies, intelligently queries Electronic Health Records (EHR) to gather clinical evidence, and prepares a comprehensive briefing for human Medical Directors.
It acts as an intelligent pipeline for:
- Parsing dynamic insurance policies via RAG.
- Querying EHR notes, pharmacy logs, and imaging histories.
- Synthesizing and verifying clinical criteria.
- Compiling deep-linked evidence packets for fast human approval.
Built with a modern, cyclical ReAct architecture featuring LangGraph.js, structured tool-calling, and strict infinite-loop prevention.
- ReAct Agent Loop (Observe, Think, Act, Reflect) → The agent dynamically adjusts its search strategy (e.g., if "Physical Therapy" fails, it searches for "NSAIDs" or "Chiropractor").
- Strict Loop Bounding → Hard limits on iterations prevent runaway compute costs and infinite loops.
- Adversarial Resilience → Built to handle conflicting medical evidence (e.g., conflicting doctor notes) by gracefully degrading and requesting human review instead of guessing.
- Medical Director Dashboard → A clean, Next.js UI for the Human-in-the-Loop to review, approve, deny, or request more info.
- Reasoning Trace Viewer → Complete transparency into the agent's internal monologue, tool arguments, and raw results.
- Dynamic Policy Retrieval (RAG) → Automatically loads the exact payer rules based on the patient's insurance and requested procedure code.
- State Checkpointing → Long-running graphs can be paused, persisted to a database, and resumed.
- ⚛️ React + Next.js (App Router) → Full-stack framework + TypeScript
- 🎨 TailwindCSS → Utility-first modern styling
- 🤖 LangGraph.js → Core orchestration, cyclic state machines, and hitl control flow
- 🧠 AzureOpenAI GPT-5 → LLM reasoning and structured tool calling
- 🛡️ Zod → Strict schema validation for all agent inputs/outputs
- 🛢️ PostgreSQL / MemorySaver → Agent state persistence and checkpointing
- 🔍 Vector DB (pgvector) → RAG for dynamic medical policy retrieval
- 🚀 Vercel → Production deployment
Feature Categories → Directories
/src/lib/agent/→ LangGraph state, nodes, conditional edges, and tools/src/lib/mocks/→ Fake patient data, FHIR simulators, and policy rules/src/app/api/prior-auth/→ Next.js Route Handlers invoking the agent/src/components/→ Dashboard UI, Reasoning Trace Viewer, Evidence Cards
1. Request→ Payload containing Patient ID and Procedure Code.2. Pre-fetch→ Backend pulls relevant policy via RAG.3. Graph Run→ LangGraph executes the Reason → Act → Reflect loop.4. Handoff→ Graph pauses; structured output is sent to UI for human review.
- Designing Cyclic State Machines (Directed Cyclic Graphs) using LangGraph.js instead of fragile, linear LangChain pipelines.
- Forcing LLMs to return strict JSON data structures using Zod and native structured tool calling.
- Implementing Human-in-the-Loop (HITL) architecture and state checkpointing for high-stakes AI applications.
- Adversarial Prompt Engineering: Training an agent to recognize conflicting data and gracefully degrade rather than hallucinate.
- Bridging complex, backend AI orchestration with a clean, transparent React frontend.
- Translating messy, real-world business logic (healthcare rules) into deterministic code boundaries.
Create a .env.local file in the root directory:
OPENAI_API_KEY=
OPENAI_ENDPOINT=""
OPENAI_API_VERSION=""
OPENAI_DEPLOYEMENTNAME=""
OPENAI_EMBEDDING_DEPLOYEMENTNAME="text-embedding-3-small"
OPENAI_INSTANCE_NAME=""
DATABASE_URL="postgresql://postgres:mysecretpassword@localhost:5432/postgres"
JWT_ACCESS_SECRET="sushantsinghnegideveloper"
JWT_REFRESH_SECRET="sushantsinghnegideveloperpro"
GITHUB_SECRET=""
GITHUB_ID=""
NEXTAUTH_SECRET=your-long-random-secret
NEXTAUTH_URL=http://localhost:3000