At Qantev, I architected a claims registration pipeline for health and life insurers that combines classification, OCR, and extraction, with low-confidence fields held for human review. I cut document classification inference time by about 10x while matching or slightly beating the previous model's accuracy.
I also orchestrated a multi-agent fraud engine, grounding validated fraud signals in claim data and pharmaceutical database checks. Cost-aware model routing sent simple claims to XGBoost and reserved frontier models for complex cases.
At AI Dynamics, I directed work on NeoPulse Manager, a model lifecycle platform for cloud, on-premise, edge, and hybrid targets. Earlier, at INSSIO, I built features for VistaClaim and created a Xamarin self-service capture app for vehicle owners.
