writing

Notes from building spawn.

Working notes on routing coding tasks to the right model: the classifier under the hood, the design calls, and the reasoning behind them.

  1. engineering8 min read

    How I trained a small coding task classifier to route models in ~50ms

    An LLM judge to pick a model is slower and pricier than the work it's deciding. Here's the tiny on-device classifier I shipped instead: what it predicts, how I built the data, and why ~50ms on a CPU was the whole point.

    Read
  2. manifesto6 min read

    Why I built spawn, an on-device model router that's adaptable to your workflows

    Every coding task goes to whatever model you opened the harness with, usually the most expensive one. Switching by hand forty times a day is a thing nobody does. spawn fixes the routing, not the models.

    Read

Stop sending typos to your smartest model.

Read the source