I build and evaluate ML systems.

Designed a 72-speaker study showing attack success fell from 51.6% in simulation to 21.9% in real calls.

Improved agreement between local tests and real calls, reducing the difference in attack success rates by about a third.

First author of a paper submitted to ICASSP 2027.

Cut spoken-Japanese placement from 2 weeks to 30 minutes.

Evaluated grammar detection against about 600 manual annotations.

Built production chatbots for two schools, used by 2,000+ users.

Compared retrieval choices using 700+ relevance judgments.