field notes from the machine
A Scotsman who dabbles in development.
What a machine full of concurrent AI coding projects actually teaches you, written down honestly.
Field notes / orchestration
A compliance check said the research harness fired in half the runs. The runs it passed had quietly skipped the one step that check existed to guarantee. Invoking a thing and running it are different claims.
2026-08-17
Field notes / local models
Nine pieces of software for talking to a local model, same GPU, largely the same models, one afternoon each. The spread in outcomes was enormous, and almost none of it was about the model.
2026-08-08
Field notes / local models
My laptop reports 24GB of video memory. It has 8GB of VRAM. Four separate numbers turned out to be measuring something adjacent to what decides whether a model loads, and not one of them was a bug.
2026-08-01
Field notes / tooling
We wanted a narrator that was Scottish, natural, local, and obtainable. Nearly every candidate managed three of those and broke the fourth. What finally worked was training on the speaker's own studio recordings, which turned out to be the source of the voice we had been cloning all along.
2026-07-26
Review notes
A green check is only evidence if you know what would have made it red. Some of ours came back clean without having tested anything at all.
2026-07-25
Field notes / tooling
Our own cost dashboard flagged the day it was born as a 26.5x overspend. The transcript said otherwise. Making sure it never happened to anyone else took a bug report, and three days later the fix shipped.
2026-07-18
Field notes / orchestration
Companion to "Route, don't guess": nine tools crossed the desk during the ten-day build. Not one became a dependency. The reasons why are the actual argument for building it in-house.
2026-07-16
Field notes / orchestration
How we built the orchestrator that routes each task to the optimal model, and the near miss that drew the line on what it's allowed to act on.
2026-07-15