All blogs
Full archive of notes on research, methodology, and the engineering that makes models usable.
What models see when they say they understand
Placeholder note on probing vision–language representations — what lights up inside the model, and how far that is from a clinical explanation you can trust.
Conformal prediction is not a confidence score
Why coverage guarantees change what a model is allowed to say, and what breaks when the exchangeability assumption quietly fails on hospital data.
Hallucination is (partly) a decoding problem
Notes from PCCD and CAST on how much fabricated content you can remove at decoding time before you have to touch the weights.
Notes on getting a polyp detector to 76 FPS
Quantization, operator fusion, and the unglamorous profiling work that separates a paper number from something a clinician can actually use live.