Blog
Field notes from production: systems, tools, and the lessons between commits.
4 min read
Private Code Search: The Parameter Nobody Wrote
Private semantic code search with Repowise, LiteLLM and Ollama: embeddings on the laptop, answers from a 120B model behind the proxy, and the one field the OpenAI client sends that nobody wrote.
4 min read
The Missing Gauge
Claude Code knows exactly how full your context window is and never shows you. A status line in under a hundred lines of Python puts the number on screen, and auto-compact stops being a surprise.
3 min read
The Cost of a Token
A demo costs cents; a shipped feature becomes a line item someone has to own. The AI bill is not the model's price, it is how you spend tokens: warm caches, small focused sessions, and knowledge kept out of the window.
6 min read
Anatomy of an Agent Fleet
One prompt, one perfect answer is a demo. A working fleet is structure: hooks gating every tool call, deterministic orchestration, and agents that never review their own work.
3 min read
The Empty Poll
Polling asks “anything new?” a thousand times to hear “no.” How a colleague's notification stopped arriving late: Server-Sent Events, the custom-header trap, and the reconnection that actually mattered.
5 min read
The Half-Life
In 2018 I spent my evenings buying certainty, and it expired. What AI automated was never my engineering: it was the lookup table I had mistaken for engineering.
4 min read
The Rule of Three
If you type it more than three times a day, it’s not a task, it’s a tax. Aliases, shell functions, and the Node version trick that lets ten runtimes coexist without a single collision.
4 min read
The Antivirus Tax
Windows Defender quietly slows every build on your machine. Here’s the one-command fix that claws back 40% without turning anything off.