About

I am an ML architect at PayPal.

Most of my work now is agents, and specifically the harness around them: verifiable environments, programmatic graders, and authority scoped so the model proposes and deterministic code performs every write. The agents themselves were the easy part. What took the design effort was the envelope.

Before that I spent a few years on the same question in a different form. I designed the canary evaluation, shadow mode, and graduated autonomy framework that became the standard here for how much autonomy a model is granted in production. Then I built hierarchical reinforcement learning for capacity allocation across roughly 1,550 applications and 100,000 pods, which is where I learned most of what I believe about reward design, largely by watching a policy find the gap between the metric I wrote and the outcome I meant, and then watching my fix push the gap somewhere I was not looking.

Worth being precise, because this distinction gets blurred: every one of those failures was optimizer-found, not model-reasoned. The policy is small and non-linguistic. It cannot represent its own evaluation or notice that it is being tested. This is reward and grader engineering under optimization pressure, not alignment research.

Earlier: cryptographic systems and protocol design under Doug Crockford. Planetary rover localization at CMU for the Google Lunar XPRIZE effort. BERT for security-critical threat detection. Variational quantum algorithms with IBM. Graduate research on how much to trust a judge and how to combine judges who disagree, which won AAAI HCOMP’s Test of Time Award in 2024, eleven years on. That problem is now called annotation quality and LLM-as-judge reliability; the benchmark is still up at github.com/utir/square.

Five peer-reviewed publications; 362 citations across all work, one paper cited 267 times; one issued patent and one pending application as lead inventor. Peer reviewer for Information Processing & Management and Artificial Intelligence.

This blog is where I write about what happens when automated judgment meets production. Mostly the parts that did not work.

Find me on LinkedIn, GitHub, or Google Scholar. My résumé is here.