R&D
My workbench for language, data & the record. Where I take words, names, places, the world's languages, prices, and the long memory of the open web, and turn them into things you can look at.
your rag eval number is hiding regressions
ragdiff: a small open-source tool that diffs two retriever configs per-query instead of scoring one in aggregate — because a metric that went up can still be hiding queries that got worse.
does the knowledge graph earn its keep?
benchmarking graph-augmented retrieval against pure vector RAG for the "what should this student see next?" problem, and auditing my own numbers until they held.