PolicySync - Clinical Policy Intelligence

Life is too short to read 40 pages of policy language just to answer one coverage question. PolicySync turns that manual grind into fast, searchable, AI-powered insights.

Comment

About the Project

PolicySync was inspired by a real pain point in healthcare operations: policy intelligence is still largely manual. Teams spend hours reading long payer PDFs, comparing policy language across insurers, and trying not to miss small changes that can impact access, compliance, and contracting decisions. We wanted to turn that process from document-hunting into decision-ready intelligence.

How We Built It

We built PolicySync as an end-to-end AI workflow, not just a chatbot on top of documents.

  • Frontend + API: Next.js (App Router) with TypeScript
  • Authentication + Roles: Auth0-based access control
  • Database + Vector Search + Storage: Supabase (Postgres + pgvector + Storage)
  • Ingestion + Extraction Service: FastAPI + LangGraph pipeline
  • AI Stack: Groq/Cerebras-powered extraction and answer synthesis with embedding-based retrieval

Core Flow

  1. Ingest policy artifacts from payer sources
  2. Detect updates using content hashing
  3. Extract and chunk policy text into structured data
  4. Route documents to payer-specific extraction logic
  5. Evaluate extraction quality
  6. Send extraction output through a human review gate
  7. Publish approved rules to power search, Q&A, comparison, and change log features

Challenges We Faced

1. Payer document heterogeneity

Different payers organize policies in very different formats, so a one-size-fits-all parser was unreliable. We addressed this with payer-aware extraction paths and fallback logic.

2. Trust and governance

In healthcare workflows, full automation without oversight is risky. We implemented a human-in-the-loop review stage and audit logging so AI supports decisions without removing accountability.

3. Reliability under constraints

We had to handle rate limits, provider fallbacks, and retrieval quality under hackathon timelines while keeping the product demo-stable.

What We Learned

This project taught us that successful AI systems depend on workflow design as much as model quality. The biggest gains came from combining:

  • better retrieval (contextual chunking + HyDE-style query expansion),
  • human review and governance,
  • and clean separation between draft and published intelligence.

We also learned how to coordinate a full-stack, multi-service architecture under time pressure, from auth and infra to extraction pipelines and product UX.

Built With

  • auth0
  • beautiful-soup
  • cerebras-api
  • fastapi
  • groq
  • hugging-face-inference-api
  • langgraph
  • next.js-16
  • next.js-api-routes
  • pdfplumber
  • pgvector
  • python
  • python-docx
  • railway
  • react-19
  • sentence-transformers
  • sql
  • supabase-postgresql
  • supabase-storage
  • tailwind-css
  • typescript
  • vercel
Share this project:

Updates