AI Engineering

Two production bugs from a legal AI platform, and why the fix for a non-deterministic LLM is engineering the boundary it reads, not writing a better prompt.

The architecture I use for agent-powered tools that run on the user's own machine and credential, with the file system as the contract and no vendor lock, drawn from an open source tool called Marketing Ops.

How I built the multi-tenant MCP server for PI6: what I expose to an agent, the OAuth security around it, and how I dogfood it with an LLM.

Notes from daily work with Claude Code on what genuinely gets faster, and where deciding what to build still takes experience.