NewMed Denial-Prevention Co-Pilot

US medical practices submit claims to payers and get 10-15% denied. Each denial costs $25-118 of labour to decode, correct, and resubmit, and about half are written off. The industry's default answer is denial management, working denials after the payer says no. But most denial reasons (stale eligibility, missing prior auth, missing modifiers, timely-filing breaches) are knowable before the claim is ever submitted.

AI product owner across the whole chain: SQL discovery, model, evaluation, and the rollout plan. Built solo as a case study for how I believe AI products should be built.

• ▸Earned the right to build: eight SQL discovery queries decided whether the problem was worth a model at all (15.4% denial rate, $72.9k of rework in six months, 80.1% of denials traceable to pre-submission causes, no self-healing trend)
• ▸Chose an interpretable 15-weight logistic model over a black box, so a biller can see why a claim is risky and trust the fix
• ▸Evaluated honestly with a temporal backtest (AUC 0.853), then turned the operating threshold into a product decision: a slider where precision, recall, workload, and dollars protected update live

• ▸A five-tab prototype scored live in the browser: the SQL discovery trail, the claim lifecycle, the working pre-submission queue, the evals and rollout plan, and a plain-English glossary
• ▸A biller's queue where clicking a claim shows its risk score, the contributing factors with model weights, the predicted CARC denial code, and one-click fixes that re-score in real time

• ▸Demonstrates the full arc, from is-this-worth-building through interpretable model to human-in-the-loop rollout, not just a model in a notebook
• ▸Every override is logged as a training label, so the system is designed to improve from the biller's judgement