About

I build evaluation and systems infrastructure for tool-using agents. At Google Cloud AI, I work on agent framework efficiency and quality, with a focus on making systems faster, more reliable, and less expensive.

Email me GitHub LinkedIn

Selected work

all research

Semantic trace summaries

Can summaries of state-changing tool calls help predict whether an agent succeeds? My study of public τ-bench trajectories found no reliable overall gain when both the task and agent were held out.

Read the study

Calibration scores and model-by-domain effects for semantic trace summaries. — Post-hoc calibration and variation across held-out models and domains.

Strust

I founded Strust to measure code-conversion performance and conformance in legacy modernization.

Read the project case study

Latest writing

all writing

Jul 9, 20264 min read (estimated)

Why p99 Increases While Average CPU Utilization Stays Flat

A request can accumulate queueing delay that average CPU utilization does not reveal, so measure its path before increasing the worker count.

Education

Columbia University: MS in Machine Learning and Bioinformatics.
UC Santa Barbara: BS in Chemical Engineering.
UC Santa Barbara: BS in Statistics and Data Science.

Other interests

I'm especially interested in agent evaluation, ML inference, GPU systems, compilers, and runtimes. I also publish reproducible studies on reliable software.

Outside of work I play cello, backpack, write, and release jazz and blues recordings as Copperbelt. I also share photographs from my travels.