AI + HEALTHCARE
Upcode Buster.
Forensic AI audit for medical billing
Single claims, batches, URLs and files through the Hono API.
Billing errors hide in volume.
Billing errors hide in volume: a procedure coded higher than the diagnosis supports, one service split into several charges, care billed that never happened. Reviewing claims one by one does not scale.
A semantic audit, then a map of the patterns.
Each claim goes through a semantic audit with Gemini 3 Pro and Flash that cross-checks procedure codes (CPT) against diagnoses (ICD-10), scores the risk and drafts an appeal letter. Results are clustered with K-Means in the browser and shown in a 3D lab built with React Three Fiber, where coordinated patterns appear as clusters. A Hono API on Supabase ingests single claims, batches, URLs and files, and an MCP server exposes the same audit tools to AI agents.
Every flag comes with its reasons.
The model explains, but it does not get the last word. Every flag carries its reasoning and risk level so an auditor can verify it, and agents reach the audit through typed MCP tools instead of improvising their own.
A working pilot, claims kept to what is built.
A pilot built at Soluntech. The audit engine, the 3D lab, the ingestion API and the MCP server are implemented. No adoption, revenue or accuracy figures are claimed.
The Upcode Buster repository and its technical case notes. Code and internal links stay private. This page summarizes what those documents show.