Research + engineering
Learning methods, scalable generation, and operational ML. Each project connects a concrete problem to a contribution and a measured result.
Language agents / Learning
Learning action scores from the states an agent actually reaches, rather than relying only on offline training data.
Foundation models / Adaptation
Diagnosing unreliable layers and adjusting how model updates are combined—without task examples or validation search.
ML systems / Distributed generation
An asynchronous pipeline separating model generation, token-level scoring, and evaluation across GPUs.
Applied ML / Operational forecasting
Designing and maintaining UConn’s weather-to-outage forecasting workflow, from data readiness through automated delivery and recovery.
SIGMA aligns different strings representing the same molecule, improving generation and prediction without adding inference cost.
For smaller systems projects and implementation notes, see Engineering notes.