what i'm up to

building the document-classification pipeline at intain — training models and shipping the product surfaces that make asset-backed loan diligence faster.

going deeper on retrieval evaluation (reranker metrics, long-context trade-offs) and spending evenings reading about structured-output generation.

this site. writing more often, posting the rough version, iterating in public.

the brothers karamazov (slowly). a pile of papers on test-time compute. too many open tabs.