13 Days, $500M in Fraud, One Binary

We took first place in the CMS Track of the ACT-IAC 2026 AI Hackathon. The system we built — Integrity — identifies over $500M in likely Medicare fraud using nothing but publicly available data. No proprietary feeds. No claims data. No protected health information.

What Integrity Does

Integrity ingests 548 million rows of public CMS data — provider enrollment, utilization statistics, payment records, geographic patterns, exclusion lists — and scores 1.47 million providers on fraud risk. It combines statistical anomaly detection with graph-based relationship analysis to surface providers whose billing patterns, peer networks, and geographic footprints match known fraud signatures.

The key numbers: 42% Precision@100, meaning 42 of the top 100 flagged providers are confirmed fraudulent. That's an 890x lift over random selection. More importantly, Integrity flags these providers an average of 6.9 years before they appear on the OIG exclusion list. Years of fraudulent billing, caught earlier.

One Binary, Zero Excuses

Integrity is a single compiled Go binary. SQLite for storage. No Kubernetes. No Spark cluster. No cloud-managed anything. It runs on a single server for under $800/month. The entire pipeline — data ingestion, feature engineering, graph analysis, scoring, and a web UI — deploys with two commands.

This wasn't a constraint. It was the strategy. We've written about why most government systems are overbuilt by 100x. Integrity is what happens when you take that philosophy and apply it to a hard analytical problem. The simplicity isn't a tradeoff — it's what made 13 days possible. No time lost configuring infrastructure. No time lost debugging distributed systems. Every hour went into the actual problem: finding fraud.

GraphWizard

Fraud is a network problem. Providers share addresses, billing patterns, referral chains, and ownership structures. To analyze these relationships at scale, we needed graph algorithms — PageRank, community detection, centrality measures — but we didn't need a graph database. So we built one.

GraphWizard is a Go library with 40+ graph algorithms that operates on in-memory structures. No query language. No server process. No network hops. Just function calls. It processed Integrity's entire provider network — millions of edges — in seconds on a single core. We open-sourced it after the hackathon. You can find it on our open source page.

13 Days

The hackathon ran 13 days. In that window, a two-person team built a production-grade fraud detection system that outperforms approaches requiring orders of magnitude more infrastructure and budget. Not because we worked harder, but because we refused to solve problems we didn't have.