I like AI projects that have a big impact, especially for clinicians. I trained as a veterinarian and I build software, so I have usually seen both the clinical side and the data side of whatever the problem is. Through my consulting firm, Datatroph, I take on fractional Chief AI Officer roles and technical due diligence, so if that's what you need, or you'd just like to talk, email me.
A clinical assistant over decades of VIN's clinical knowledge that cites the source of every answer. I built the first version, an engineering team runs it now, and practicing vets use it every day. VIN itself reaches more than half the vets in the US. Knowing when it is wrong is harder than building it.
With the canine genetics group at NC State, applying NLP and machine learning to paired clinical and genomic data. It is early work.
Clinical AI is being scored against a ground truth that’s less solid than it looks, and I coauthored a paper on why that is and what to do instead (arXiv, 2026).
Read the paper →A research graph for Claude Code that keeps track of what’s been tried and what held up. I wrote it for my own work and it is public in case it is useful.
On GitHub →PLOS ONE, 2020
A population-level picture of antibiotic use in Australian companion-animal practice, pulled from free text rather than surveys.
BioNLP Workshop @ ACL 2020
VetBERT. A domain-adapted language model plus instance selection cuts annotation effort enough to make clinical text classification practical.
JAC-Antimicrobial Resistance, 2022
Whether prescribing matched guidelines, at dose level, across millions of records; the kind of question you can only answer with the notes.
arXiv, 2026
Why evaluating medical AI is harder than it looks when the ground truth itself is uncertain, and one way to reason about it anyway.