Notes from the
build floor.

A field report from a dozen RAG deployments. The failure modes are remarkably consistent, and most of them are fixable in an afternoon.

A walkthrough of the eval harness we attach to every AI project before model selection, prompt design, or anything else interesting.

pgvector vs Pinecone vs Weaviate vs Turbopuffer, what we use, when, and what the gotchas look like at million-scale.

Guardrails, tool-use schemas, and the eval suite we run before any agent gets production credentials.

The user interface for an AI product is half text, half waiting. Lessons from a year of shipping conversational and copilot UI.