Writing

Two series, both bottom-up: the scaffolding around a language model, and the physics underneath a speech recogniser. In each case I'm less interested in the happy path than in the layer below it — what the abstraction is standing on.

Harness engineering is about what surrounds a model in production: agent loops, tool schemas, and the failure modes that only appear once something is running.

Speech from first principles derives the acoustic model of the human voice from the wave equation, then runs it backwards to show where every piece of a speech recognition front end came from. Both posts have diagrams you can drag and sounds you can play, and the numbers in them come out of code that ships with a test.

• Running the tube backwards

  11 August 2026 · Speech from first principles · Part 2 · 16 min read

  The acoustic model of a speech recogniser, derived as the inverse of the one that makes the sound.

• How a tube becomes a vowel

  11 August 2026 · Speech from first principles · Part 1 · 20 min read

  Deriving the acoustic model of the human voice from the wave equation — with every step you can hear and take apart.

• Your tool exited 0. It did nothing.

  21 July 2026 · Harness engineering · Part 1 · 8 min read

  What a misbehaving RGB controller taught me about designing tools for language models.

Subscribe via RSS