We build and study machine learning systems across vision, language, and medicine, uncovering how they succeed, how they fail, and where their confidence can't be trusted.
We're a small team with backgrounds from Johns Hopkins, Cornell, and Rutgers University.
Most published results describe what a system gets right. We spend most of our effort on what it gets wrong: which classes, which patients, which formatting conventions, which prompts, and whether the system knew it was wrong when it mattered.
We publish research on robustness, calibration, and failure structure across vision, language, and medicine, an ongoing, growing body of work, released with code, data, and model weights so every claim can be checked.
01 · Audit before we build
Before we improve a system, we measure exactly where and why it fails, at the class level, the patient level, and the confidence level.
02 · Replicate across conditions
A pattern only counts if it survives independent datasets, different models, and different populations, not a single benchmark.
03 · Release everything
Code, data, and model weights ship alongside every paper, so our conclusions can be checked, extended, or broken by anyone.