Billing integrity pipeline

The primary billing integrity system: it checks every timesheet submission, every weekday. An LLM classifies the free text of each note, statistical baselines compare it to the population and to the provider’s own history, and a readable rules matrix decides what gets escalated to a person. It’s a guardrail, not a ranking.

I built it, from the reference implementation to the dbt pipeline, chose its model through blind adjudication over the full history, and wrote the spec for the real-time version that catches problems before submission. I did the data science myself and oversaw the engineering for the production migration.

More in AI doesn’t have bad days.