Selected Work
Source linked where public.
-
SREGym
Training environment for on-call agents. Each seed breaks a small production stack a different way, and a verifier checks whether the agent actually fixed it.
-
GUARD
Watches a PyTorch model in production for drift. Runs on the GPU and needs no labels.
-
mimir
Runs prompt and model experiments and tells you whether the difference is real or just sampling noise.
-
Origin
Tells you whether text was written by a person or a model, and which sentences it thinks are which.
-
Relay
Catches agent actions mid-execution, blocks the bad ones, and rolls back what got through.
-
Latent Actions from Factorized Transition Effects under Agent Ambiguity
Nam, Jonnalagadda, Aggarwal, Xu, Balestriero. ICML Workshop on Compositional Learning, 2026. Learning an agent’s actions from video breaks down when the camera and background move too. Splits each transition into reusable pieces first, then builds the action from those.
-
In Silico Identification and In Vitro Validation of Repurposed Compounds Targeting the RSV Polymerase
Xu, Park, Calderon, Cao, Liang. Microorganisms 11(6), 2023. Screened approved drugs for ones that block the RSV polymerase; Micafungin worked in vitro.