Lebovsky

## Accomplishments that we're proud of Built a full end-to-end MVP instead of just a mock UI or isolated model demo. Turned unstructured payer policy documents into structured, queryable coverage data. Created an evidence-backed extraction flow where every usable field must be supported by source text. Added policy versioning and structured change tracking, not just point-in-time search. Built a clean dashboard for search, compare, evidence review, and policy ingestion. Designed the system so it works with both mock extraction and real LLM providers like Vertex AI.

What we learned

The hardest part of this problem is not calling an LLM, it is building a trustworthy system around it. Evidence and validation matter more than raw model output quality. Policy documents are messy, repetitive, and inconsistent, so normalization is a major part of the product. Versioning and diff logic are critical if you want the product to feel useful for real policy monitoring. For healthcare workflows, a structured interface with traceability is more valuable than a chatbot-style experience.

What's next for Lebovsky

Ingest more real payer documents from the starter packet and beyond. Expand extraction coverage across more drugs and policy formats. Add a stronger analyst review workflow for low-confidence or partial extractions. Support scheduled monitoring so new policy versions can be detected automatically. Improve production readiness with better deployment, logging, and operational monitoring.

Built With

• fastapi
• python
• react
• sqlite
• typescript
• vertex
• vite