Policy-to-Rules Engine

Overview

A proof of concept built for a healthcare payment-integrity assessment. It takes a written payer medical policy (prose) and converts it into structured, machine-readable claim-editing rules using an LLM, then runs sample claims against those rules and returns approve/deny decisions — each with a reason tied back to the source policy.

This maps directly to how payment-integrity products convert written clinical and payment policy into executable claim edits.

Problem

Payers write coverage and payment policy in English. Turning that prose into executable claim edits is manual, slow, and error-prone — and when a claim is denied, the reasoning often isn't traceable back to the specific policy language that drove it.

Solution

A deliberately simple two-agent pipeline that demonstrates the core payment-policy-management loop end to end:

• Extraction agent reads a written policy and emits a structured rules JSON object against a defined schema
• Validation agent takes those extracted rules plus a single claim and returns a decision — APPROVE or DENY — with the matched rule and a source-policy reference

The two-agent split shows an agentic design without unnecessary complexity: no database, no auth, no extra services. Every decision is traceable to a source line.

Architecture

Flow

1. Written policy (prose) → Extraction agent → structured rules JSON
2. Rules JSON + one claim → Validation agent → { decision, matched_rule, policy_reference }
3. UI renders each claim as an APPROVE / DENY card with the reason and the policy line it maps to

Design choices

• Robust JSON handling — both agents are instructed to return raw JSON only; the pipeline still defensively strips code fences and wraps parsing in try/catch
• Traceability first — every adjudication points back to the exact rule and source policy reference
• Screenshare-optimized UI — a clean three-column layout (raw policy → extracted rules → claim results) built for a live demo

Tech Stack

• Backend: Node.js + Express (thin API layer)
• AI: OpenAI SDK with a structured extraction prompt + schema and a separate validation prompt
• Frontend: Plain HTML + vanilla JS single page — three columns, no build step
• Config: API key from .env, never hardcoded (only .env.example committed)

Demo Walkthrough

1. Show the prose policy on the left — payers write coverage policy in English
2. Click Extract — an LLM agent converts prose into machine-readable claim edits
3. Click Run all claims — walk through the results: a clean approve, denials on diagnosis / frequency / wrong CPT, and an approve under a malabsorption exception

Results

• Demonstrates the first principle behind automated payment policy management: written policy → executable rules → adjudicated claims
• Every decision is traceable to a source policy line
• Delivered with a written report and presentation alongside the working demo