Currently
Now
Updated May 2026 · what is this?
A snapshot of what I'm focused on this quarter — research, code, reading, life. The page is intentionally short and changes a few times a year; for anything more durable, see publications, projects, or writing.
Working on
- Model understanding for LLM agents at Meta — attribution workflows, prompt & context optimization, large-scale ML debugging. The interesting frontier right now is making attribution useful as an intervention, not just a visualization.
- Surrogate fidelity — when can open LLMs faithfully explain closed ones? With collaborators at Columbia and Meta. Just landed as a Spotlight at ICML 2026 Mech-Interp Workshop.
- PyTorch Captum — keeping the interpretability library healthy. Release engineering, GenAI attribution support, type/test cleanup, and the occasional new method.
Recently landed
- This site. A long-overdue rebuild on Astro + Tailwind, replacing a Django/Bootstrap site that hadn't moved much since the PhD.
- Animated paper diagrams for Unfooling, ProtoFlow, and PixPNet — explaining each method's core mechanism in three CSS-animated acts.
- AIware 2026 paper on hierarchical hunk-level attribution for LLM-based code-risk models — the production story behind a year of attribution work.
On my mind
(Coming in the next refresh — questions I'm chewing on right now.)
Reading
(Coming in the next refresh — papers, books, essays worth pointing at.)
Last updated . The /now page convention is from Derek Sivers.