Building AI systems that have to hold up

I build production AI systems for healthcare and pharmacy intelligence — systems that have to hold up against dense clinical literature, drug data, and guidelines where being approximately right is not good enough.

Clinical and pharmaceutical information is a genuinely hard substrate for AI. It is dense, fragmented across sources, frequently contested, constantly updated, and read by people making decisions that matter. A system that is fluent but unreliable is worse than useless in that setting — so most of the engineering goes into grounding, verification and measurement rather than generation.

My work spans the full lifecycle: architecture and experimentation, then evaluation, then production deployment and everything that keeps a system correct once real users depend on it. Prototyping against a language model is the easy part of that arc, and the least interesting.

Core Sports AI turns match footage into performance intelligence. Computer vision reads the match — players, ball, shots, positions — and the analytics layer turns that into the kind of read a good coach gives you, at a scale no one can watch by hand.

I lead technology and product end-to-end: the AI architecture, the computer vision stack, the analytics systems, the product itself, and the technical direction of the company.

SOLS explored whether AI could predict infrastructure defect events before they surfaced, give operations and maintenance workers concrete resolution guidance, and let teams interrogate predicted maintenance information directly instead of reading reports about it.

SOLS was an early lesson in what applied AI actually demands outside a research setting: messy operational data, users who need an answer they can act on rather than a probability, and a domain where the cost of a wrong prediction is measured in site visits and downtime. It shaped how I think about deploying AI into real workflows.