Radiology Specialist · Chief Innovation Officer · Built CBCTScope
I build AIwhere being wrongisn’t an option:medicine.
Most AI advice has nothing at stake. I read medical images every day as an oral and maxillofacial radiologist. I build the training data from my own reads, train the model on it, and test it on cases it has never seen. These essays keep that bar: what a model did on a real case, what it did not, and how to build alone without fooling yourself.
100,000+ reports signedCBCTScope shipped, open sourcewhat I’m building →
// Latest writing
One-Person Startup5 minAgent-native imaging: why I gave my CBCT viewer to the agents
Why I built CBCTScope, an open-source local-first CBCT viewer with native AI-agent control over MCP, and why the agent gets navigation verbs but never a verb that returns a finding.
One-Person Startup10 minFine-tuning got cheap. The expert who grades it did not
Fine-tuning got cheap. Graded truth did not. The scarce input to specialist AI is a person whose corrections are correct. Five budget lines to write first.
One-Person Startup9 minSame AI model, different answer: record the runtime too
Same weights, same prompt, same seed, opposite answer weeks apart. The inference stack moved. Five lines make an AI evaluation number comparable with itself.
One-Person Startup10 minWhat not to teach an AI model from your own expert reads
A model trained on your judgment learns your floor, not your ceiling. Label only what the input supports, keep the rest, and decide it before the first example.
One-Person Startup10 minWhy you must read the image before you open the AI answer
Once you have seen the AI's answer you are checking a claim, not reading. Radiology fixed this by blinding the second reader. Five steps bring it to any review.
// Lab Notes
All entries// Browse by series
// Agent-native instruments, open source
Newsletter
Get the essays as they’re written.
New pieces land in subscribers’ inboxes first. Free, no noise.