dumb things I always wanted to write about
What production conversational AI takes once you stop treating it as request and response: instant talk, slow thinking, one hard invariant, and patience for the third of users who send the next message before you have finished answering the last one.
An Elixir agent framework built on gen_statem. Stateful processes, context aging, multi-agent coordination, and why the five-line demo disappears the moment an agent has to survive a crash mid-response.
An append-only knowledge graph that records how a person's life changes over time. Deterministic, causal, and fast enough to sit inside a real conversation.
Treating context as a living system rather than a retrieval problem: a temporal hierarchy, freshness states that expire on their own, and a scratch pad the model reasons in before it answers.
The dual-buffer design and the coherence invariant underneath it: what the user sees, what gets saved, and what the model knows have to be identical. It sounds obvious. It was not.
On the ninety-nine percent of design work that nobody screenshots, and why boring is usually the correct output.
Why I stopped reaching for Bootstrap, six months into using Tailwind on personal projects and in production.
The basics of interface design for people who build things but do not design them.
Linear regression for the first time, without the notation getting in the way.