The reference material behind the work.
This is where I put the guides I write to teach a subject properly: to myself first, then to the engineers I work with. Each one is a complete working reference — the plain-English explanation and the engineering explanation side by side, with the failure modes named. Every worked example is a reference design unless it says otherwise, and where a real system is named, what it actually runs on is stated plainly. Project-specific walkthroughs live with their projects and are listed below.
The guides
pgvector, explained twice — vector search in PostgreSQL
Reference guide · retrieval and evaluationEvery section written twice: the plain-English version and the engineering version, side by side. Embeddings as a schema contract, HNSW versus IVFFlat, the operator-class mismatch that degrades to a sequential scan, over-fetch-and-fuse ranking, and the recall@k / MRR / nDCG golden-set gates that catch recall loss before a user does. The worked examples are reference designs, not systems I have deployed.
Cerebro — a research platform, documented end to end
Architecture guide · seven pages · lives with the projectHow a research system is built so its own numbers can be trusted: point-in-time universe construction, an execution model that makes look-ahead structurally impossible, and walk-forward out-of-sample validation. The performance layer is one I rebuilt from the account and fill streams after the engine's built-in analyzer reported wrong numbers. Cerebro is my own independent project and contains no ML.
boost — the visual guide
Project documentation · lives with the projectThe reference site for boost, my open-source agent-skills CLI: what an agent skill is, how registries, taps and the lock file fit together, and how the pure-stdlib BM25 retrieval engine and its CI evaluation gates work.