Notes from the build

Blog.

Posts about the projects I build: how a thing was put together, what broke on the way, and where I stand on the choices it took.

13 min read

How I run multiple AI agents locally on a Mac. Virtually no cost.

I run several AI agents on this Mac against one MLXLM server at localhost, port 8080. OpenCode and Xcode call the tools, the model stays on the machine, and there is no API bill. Virtually no cost.

Read

12 min read

Training a 62M LLM from scratch to read search results, and tying Qwen3-0.6B

In five days I trained a 62M small language model from scratch that answers questions by reading search results, in a 70 MB file meant for phones. On 258 real Google questions it ties Qwen3-0.6B, and a bigger model trained on my own rows still scored higher, 214 to 206.

Read

9 min read

Scaling a Node.js trading system to six workers on Redis, for $126 a month

One Node.js service for trades, rules, sockets and email became six workers behind one Redis and BullMQ. Here is what it costs, what it handles, where it breaks, and what happened when I rewrote the rule engine in Rust.

Read