research
A journal paper, a workshop paper, and two more that drive my work. Each one says what it couldn't settle.
Formalizes what happens when a messaging account's history splits across two devices with no way to recombine them, and proposes CARTA, a reconstruction contract that quarantines and reports what it can't safely resolve.
Most work on making AI accessible asks how to present what a model says to different readers. It rarely asks whether the model should issue its strongest claim at all, and to whom.
Takes one prediction from a random forest model trained on the UCI Maternal Health Risk dataset and renders it three different ways depending on who's reading it: the full SHAP decomposition for a clinician, a ranked-factor card for a community health worker, and a three-lamp visual with a spoken Hindi sentence for the mother.
Can a designer measure a property of their own data before building anything and learn which classification mechanism to build? On seven public author name disambiguation datasets the answer is mostly no, for two separate reasons.