Blog
My thoughts on building AI systems that run in production - agents, retrieval, cost and the things that break.
Two Windows bugs, two merged PRs: my first contribution to Whiteboard
Whiteboard hung at start-up on my Windows 11 machine and couldn't see Claude Code. Both turned out to be Unix assumptions in Node code, and both fixes were merged the same day.
A full video production studio, run by AI from a folder of text files
Two days setting up a complete video production suite on one laptop - audit, renderers, voice pipeline, editor - with every paid service switched off and $0 spent.
My laptop could run the model - it just couldn't download it
A 4GB laptop GPU runs MiniCPM5-2B at 45 tokens a second. Getting the 2.7GB file onto the disk was the part that took all evening.
What an LLM call actually costs
Most LLM cost tracking multiplies total tokens by a single rate, and that number is confidently wrong at exactly the point where the spend is largest.
Why I deleted five sub-agents and shipped one
A supervisor agent kept mis-routing requests across five specialised sub-agents. The fix wasn't better prompting - it was deleting four of them.