Running a coding agent on a CPU
A 3B model cannot reliably produce the JSON that tool calling assumes. Change the syntax and the same model starts finishing tasks.
5 min read
- LLM
- Agents
- Local-First
Writing
Notes from things I built or broke. Mostly retrieval, security and the unglamorous parts of shipping.
I audited 138 of my own GitHub repos
A token that had been live for nine months, a secret that a path-list scan missed and a pattern scan caught, and why rotating the credential matters more than scrubbing the history.
5 min read- Security
- Tooling
Layered RAG: why vector search is the wrong tool for half your questions
Most questions about business documents are about metadata, not meaning. Embedding them anyway is how you get a confident answer that is quietly wrong.
6 min read- RAG
- Retrieval
- LLM