Healthcare Claims Adjudication System

Hybrid rules-engine + LLM system for healthcare claims processing — deterministic logic for auditable decisions, RAG-powered reasoning for edge cases.

Live at healthcare-agent-mu.vercel.app

---

The Problem

Healthcare claims adjudication is a compliance-sensitive workflow. Every decision must be auditable — "the model decided" is not a valid answer in a regulatory audit. At the same time, the rules are complex enough that a pure rules engine leaves too many edge cases unhandled.

The standard industry approach: manual review for everything the rules don't catch. That's slow, expensive, and inconsistent.

---

The Architecture

Two-tier processing:

Tier 1 — Deterministic rules engine handles ~90% of claims. Every decision produces a line-by-line audit trail. No LLM involved. Fast, explainable, auditable.

Tier 2 — RAG-augmented LLM handles edge cases the rules don't cover. ChromaDB retrieves similar historical claims as context. Every LLM decision is logged with its retrieved context and confidence score — so auditors can trace exactly why the model decided what it did.

Human reviewers see both tiers in a unified UI and can override any decision with a logged reason.

Claim input (CSV or JSON)
    ↓
Rules engine → deterministic decision + audit log
    ↓ (if ambiguous)
ChromaDB similarity search → top-k historical claims
    ↓
LLM adjudication → decision + retrieved context logged
    ↓
Review UI → human override (optional, logged)
    ↓
Final decision + full audit trail

---

Stack

Layer	Tech
Backend	FastAPI, Python
Frontend	React, Vite
Vector DB	ChromaDB (RAG)
LLM	OpenAI-compatible API / local LLMs
Storage	SQLite (claims), ChromaDB (vectors)
Deploy	Docker Compose, Vercel (frontend)

---

Why This Is Hard

1. Auditability requirement — healthcare adjudication requires every decision to be explainable. Pure LLM systems fail here. The hybrid architecture exists specifically so every decision has a traceable reason
2. RAG context quality — retrieving the wrong historical claims makes the LLM worse, not better. The ChromaDB schema and embedding strategy matter more than the LLM choice
3. Human-in-the-loop UX — reviewers aren't engineers. The UI must surface the right context (why did the system decide this?) without overwhelming the reviewer with raw model output
4. Edge case definition — deciding what goes to Tier 1 vs Tier 2 is itself a design problem. Too aggressive with rules = brittle. Too aggressive with LLM = unauditable

---

Running Locally

# Clone and set up environment
git clone https://github.com/Shubh3005/HealthcareAgent
cp .env.example .env  # add your OpenAI API key

# Run with Docker Compose
docker-compose up -d

# Frontend at localhost:3000
# Backend API at localhost:8000/docs

---

Built by Shubham Gupta · Penn State CS, Schreyer Honors College