Building. Continuing the 0 → 1 for the first agentic workflows on the FirstParty Platform. Fine-tuning schema-aware data agents that ingest, normalise and contextualise the messy unstructured stuff that hedge funds and asset managers care about. Lots of time in evaluation harnesses; lots of time deleting features that looked clever in the demo but didn't survive contact with real users.
Thinking about. Time-Series Foundation Models (Datadog et al). With a hypothesis around buyer use cases for data products, you can query history until your heart's content. But spotting the signals that you DON'T know what to look for involves mass compute across everything. TSFM's are nominally focused on forecasting — but conceptually, they'd be helpful for unexplored nowcasting. In research mode.
Learning. Reading Code: The Hidden Language of Computer Hardware and Software (Petzold), and the SICP MIT textbook (Abelson, Sussman, Sussman) — because I probably don't know as much about actual computer processes as I should. (Both were recommendations from guests on Lenny Rachitsky's podcast.) Also (re-)reading Marcus du Sautoy's Thinking Better: The Art of the Shortcut — for fun math learnings to teach Joshua (13). Trying (as ever) to stay on top of the latest from the major labs and finance houses re AI.
Not doing. Saying no as much as I should. Too much fun stuff to learn and explore, so sleep (and a few workouts) are being sacrificed. Saw Jason Strimpel recently — who said he was perpetually paranoid about 'working on the wrong thing' and wasting this moment... which I empathize with.
Outside work. My son's 7th-grade algebra learnings. Planning summers. Dungeon Crawler Carl Book 8 just dropped (in Jeff Hays' amazing audio format). Personal website refreshing.