Current Role: Tech Lead, DIGI @ University of Washington
My research examines how probabilistic language models, calibrated cross-encoder classifiers, and deterministic domain rules can be composed safely under tight compute and reliability constraints.
Training and calibrating lightweight bidirectional cross-encoders (Laya) as a pre-generative verification gate. Enforces bounded False Discovery Rate on retrieved context, detects cross-chunk contradictions in longitudinal records, and triggers deterministic abstention before token generation.
Mapping the empirical boundary between small open-weight language models operating over unstructured clinical notes and deterministic rule engines (WHO SMART Guidelines L3 CQL and verified SQL) operating over structured data.
Developing evaluation methodologies that separate model hallucinations from human rater disagreement and guideline ambiguity. Combines Prediction-Powered Inference (PPI), Script Concordance Test (SCT) scoring, ceiling-normalized inter-rater agreement, and critical omission floors.