Writing
Pieces we write when something is worked out far enough to show rather than assert — a real mechanism, a command you can run, and the parts that don't work yet stated next to the parts that do. There aren't many, on purpose.
Worked example
Most AI document tools save whatever the model extracts. We built a tax pipeline where a number can't reach the math until something other than the extractor has confirmed where it came from — and you can run the check yourself.
Read it →Most of what we'd otherwise write up lives closer to the thing it describes. The receipts are real artifacts you can download and recompute rather than read about, and the developers page is the shortest path from "explain the idea" to "run the command".