Research & Projects

Tested whether a deliberately weakened copy of an LLM can point to the reasoning steps that decide the answer. Three interventions say it cannot.

Read transcription factor binding out of a DNA language model without labels, and showed a cross-modal model was genuinely using its protein input.

Showed that HealthBench rankings move with the prompt, and that swapping the judge model shifts scores far more than rewording the judge instruction.

Traced representation collapse in deep GNNs to optimization rather than to message passing.

Built a staged natural-language interface to an industrial database, then benchmarked models inside the fixed harness to tell model quality apart from system design.

Built a full-stack assistant that suggests outfits from a user's own wardrobe and renders them with virtual try-on.

Recovered 3D facial geometry from a single photograph by fitting parametric face models, with every stage of the pipeline separately verified.