Case Study
ClaimGuard-AI
Scores healthcare claims for denial risk before submission. Gemma extracts the fields, XGBoost scores them, a DuckDB knapsack orders the review queue.
Executive Summary
ClaimGuard-AI is a pre-submission risk engine for healthcare revenue cycle management, built for the AIxBio Hackathon. It replaces static rule-based scrubbers with three layers: extraction, scoring, and financial optimization.
Problem & Constraints
Claims get denied over documentation gaps, procedure mismatches, and payer policy violations. Manual scrubbing is slow and ignores expected financial loss when ordering the queue.
Architecture
Physician Note + CPT/ICD/Payer → Nebius Gemma JSON extraction → XGBoost denial probability → Expected Loss EL = V × P → DuckDB knapsack sort → Next.js auditor worklist.
Methodology
- Nebius Token Factory (Gemma 3 27B) extracts strict JSON clinical flags (Pydantic strict=True)
- XGBoost classifier predicts denial probability with scale_pos_weight for imbalance
- DuckDB analytical sort with bounded knapsack (capacity K/day) prioritizes by expected financial loss
- Next.js frontend: Dashboard, Claims Queue, Agent Studio, Reports with PDF/CSV export
Results & Metrics
| Layer | Technology |
|---|---|
| Extraction | Nebius Gemma 3 27B |
| Scoring | XGBoost |
| Optimization | DuckDB + Knapsack |
| Frontend | Next.js 16 multi-page |
Tech Stack
Next.js 16, FastAPI, XGBoost, DuckDB, Nebius API, Pydantic
Future Work
90-day regional health system pilot, Supabase persistence, EHR FHIR R4 integration, NVIDIA Global Sprint deployment.