Blog
Short tutorials on how I built a few things, with the code that matters.
- Claude Code + Jev: the agent writes, Jev decidesJev can't run in place of Claude Code's model. What you can do is teach the agent to write correct Jev calls, and that's what this post shows.
- Hybrid search in a single SQLite fileI put vector and text search in the same Turso database, with both queries in a single round trip. In the end, the benchmark showed keyword search wasn't adding recall.
- Turso FTS5 in RAG without 500sPass the user's question straight into FTS5's MATCH and sooner or later you get a 500. Here's how I build the query, and what corrupts the index on write.
- Weighted RRF and the vector_top_k trapWith plain RRF, an exact-phrase match barely moves a result up. Why I gave that search a weight of 2, and the vector_top_k filter that kept returning an empty list.
- An AI code-review bot on Lambda and LiteLLMA Bitbucket webhook calls a Lambda, which sends the diff to a model through LiteLLM. The hard parts were never reviewing the same commit twice and a 300-second timeout that came from fetch.
- An SQS worker for heavy work in a Next.js app on SSTIndexing a document takes minutes, so I moved it to an SQS queue with a Lambda, in the same sst.config as the site. Most of the work was getting the timeout, the DLQ and arm64 right.
