Notes
Things worth writing down
Mostly about the parts of applied AI that are not the model: retrieval quality, evaluation, and the unglamorous data work that decides whether any of it survives contact with production.
- 2026-06-14The human gate wasn't a safety feature. It was the eval set.We added approval gates to an incident agent because nobody would let it touch production. The gate turned out to be the only reason we could measure whether it worked.agentsevaluationLangGraph4 min read
- 2026-04-22Your service topology is a graph. Stop embedding it.Chunking a dependency map into a vector store throws away the one property that makes it useful. For incident retrieval, traverse first and embed second.RAGGraphRAGretrieval3 min read
- 2026-02-09Athena wasn't slow. We just gave it four million files.A small-file problem looks like a scale problem and is almost always a layout problem. One compaction job and one tunable parameter cut storage cost 40%.data engineeringPySparkAWS3 min read