🛡️ My Neighbour's ₹2 Lakh Insurance Claim Was Rejected. I Built an AI to Make Sure It Never Happens Again.

#devchallenge #githubchallenge #webdev #ai

GitHub “Finish-Up-A-Thon” Challenge Submission

This is a submission for the GitHub Finish-Up-A-Thon Challenge

What I Built

PolicyEye (originally started as SecureShield) is a GenAI-powered health insurance claim eligibility engine built for Indian patients. You upload your policy PDF, enter your case details, and a 5-agent AI pipeline tells you exactly whether your claim is eligible - citing the specific IRDAI 2024 regulation behind every decision. No guesswork. No hallucinations. No rejected claims due to ignorance.

🔗 Try the Live App: policyeye.app
🔗 GitHub

Tech Stack:

Frontend → Next.js 16 (React 19, Turbopack) Backend → FastAPI + LangGraph Database → Supabase (PostgreSQL + pgvector) Auth → Supabase JWT AI Gateway → Cloudflare AI Gateway LLM Chains → Cerebras, Groq, Gemini, SambaNova, HuggingFace, xAI, OpenRouter Compliance → IRDAI Health Insurance Regulations 2024

Demo

The Comeback Story

In 2024, I watched a friend's father's ₹2L surgery claim get rejected - because of a clause buried in page 34 of a 47-page policy nobody had read. That experience planted the seed.

Early 2025 - ET GenAI Hackathon: I pitched PolicyEye. The idea was validated by the judges. I was motivated. Then came university exams. I had to choose - submit a half-baked prototype or protect my grades. I chose my grades. PolicyEye sat abandoned: 2 rough Python scripts, local SQLite, no UI, no pipeline.

June 2026 - GitHub Finish-Up-A-Thon: The challenge felt written for me. "Finally finish what you started." So I did.

Before → After

	Before (ET GenAI)	After (Finish-Up-A-Thon)
Agents	2 rough Python scripts	5-agent LangGraph pipeline, 18 tools
Database	Local SQLite	Supabase PostgreSQL + pgvector
UI	None (terminal only)	Premium Next.js 16 frontend
AI	Single provider	Multi-provider fallback chain (7 providers)
Deployment	Ran only on my machine	Containerized, Live on Vercel & Hugging Face
Security	None	JWT, rate limiting, HMAC-SHA256
Compliance	Hardcoded rules	Semantic search over 49 IRDAI 2024 chunks

Key design choice: The Decision Engine uses zero LLMs - all eligibility verdicts are deterministic and auditable. LLMs only generate the plain-language explanation. No hallucinations in decisions. Ever.

Copilot helped me in three meaningful ways:

My Very First Deployment! - I had never deployed a full-stack application before. Copilot walked me through containerization, Hugging Face Spaces backend setup, and deploying the frontend to Vercel. Bringing this complex architecture live to the internet was truly only possible because of Copilot's guidance.
Stack decisions - Copilot helped me choose pgvector-in-Supabase over a separate vector DB: fewer services, tighter auth, same SQL interface I already knew.
CI/CD & Architecture - As the project grew (backend + frontend + 7 AI providers), Copilot generated the routing logic to automatically failover when free APIs hit rate limits, ensuring the app never goes down.

My Personal Note

I'm an engineering student in India. Insurance confusion and claim rejections are real around me, every day. AI in healthcare isn't about replacing doctors. It's about making the system legible to the people it's supposed to serve.

Thanks to this challenge, PolicyEye is finally out of my laptop and into the world.

Built with LangGraph, FastAPI, Next.js, Supabase, and a lot of frustration turned into motivation.