Research Portfolio
One Question, Two Directions.
AI can produce a slide deck, a 3-D model, or a day's worth of edits in one pass, but whether the result is good, safe, and true to what the person meant still takes human judgment. My work approaches that judgment from two directions: measuring whether AI can share it, and building interfaces that keep it in human hands.
Can AI Judge Its Own Work?
Benchmarks & Taxonomies for AI Evaluation
Datasets, taxonomies, and benchmarks that test how well models understand design quality, the consequences of UI actions, and personal context, and that make them measurably sharper when explicit structure is supplied.
Beyond Visual Defaults
Human-AI Creation Without Sight
Interactive systems through which blind and low-vision people author slides, charts, and 3-D models: perceiving layout without vision, verifying AI output they cannot see, and automating routine work without giving up control.