Writing
Notes on training, evaluating, and shipping large models, here and elsewhere. RSS.
-
Mar 2026Building AI agents for 131 million customers
Building Nubank · with Daniel Braithwaite
Five lessons from putting customer-support agents into production: why evaluation comes first, what transactional NPS actually measures, and how prompt optimization and composite tools behave at scale.
-
Sep 2025Muon for improved foundation model pretraining data efficiency
Building Nubank · with Daniel Braithwaite, Arissa Yoshida and Rafael Celente
What the Muon optimizer buys you when pretraining transaction-data foundation models, and the groundwork for quantizing its optimizer state.
-
Sep 2020GDMix: A deep ranking personalization framework
LinkedIn Engineering · with Jun Shi, Chengming Jiang and others
The open-source framework behind per-member personalization of LinkedIn's ranking models, and how generalized mixed models train at that scale.
How this site is built is documented in the colophon.