Writing
Notes on AI infrastructure and distributed systems, from production, not from a whiteboard.
3 min read
Request counting assumes requests are roughly interchangeable. LLM calls are not, one can cost ten thousand times more than another, and you don't know the price until it's over.
3 min read
The hardest part of agent infrastructure isn't the models, it's keeping track of who a request belongs to as it hops between agents and tools that would all rather forget.
2 min read
Eight years in, the most senior skill I can point to isn't a technology. It's knowing when not to be interesting.
1 min read
What I plan to write about here, infrastructure for AI agents, distributed systems, and the lessons that only show up in production.
Also on this site
22 lessons published, one a day, read or listen.