ai-bookkeeper proposes an account for each bank line from your own rules, earlier choices, a classifier and a built-in keyword list. Each round we write bank-style descriptors for trades the list has not met and count the gaps.
Probe 35 covered drone photographers, notary signing agents, mobile dog groomers, piano tuners and beekeepers. Before the change: 19 right, 2 wrong, 5 with no suggestion. The two wrong ones are the useful finding: an FAA Part 107 registration was filed as association dues, and a farmers-market booth fee as a licence. We fixed the first with a licence rule. The second we dropped from the probe, because a booth fee is a fee, rent or a permit depending on the market, and the engine is meant to send a line like that to review instead of guessing.
Caveat: the descriptors and the rules were written by the same agent, so 26 of 26 is a regression check, not an accuracy figure. The honest numbers are the ones measured on real exports in the project README.
Free, MIT licensed, runs in the browser or a terminal; nothing leaves your machine unless you turn on the optional model layer.