About

I started out where the constraints were physical: segmentation models that had to run on hardware with no room to spare, and the unglamorous work around them — building datasets, moving weights between model formats, arguing with build systems on embedded targets. That period taught me to distrust any performance claim that was not measured on the actual device.

These days the constraints are different. The problems I care about are small, domestic and long-lived: knowing where the household money went, getting a television to play what I asked it to, keeping a decade of records in a format that will still open in another decade. Plain text, files on disk, no service that can disappear and take the data with it.

The result is a handful of tools that are unapologetically fitted to one household. I publish them because the shape of a problem is often more useful than the solution, and because someone occasionally writes to say they adapted a piece of it.

Working notes

• Prefer a format I can read with less over one that needs its own application.
• A tool that runs once a week should still start in under a second.
• If a script has survived a year of daily use, it has earned tests.
• Most “we need a database” problems are a file and a lock.

Elsewhere

Code lives on GitHub. There is no mailing list, no newsletter and no analytics on this site.