ai-bookkeeper suggests an account for a bank line with rules, payee memory, a naive Bayes model and built-in keywords. We measured the keywords on card-charge memos from Hack Club's public HCB feed, organisations on pages 2 and 3 of the list, which no rule had been written from: 560 charges.
Before: 218 suggested (38.9%). Memos carry no category, so this is coverage, not accuracy. The misses were bare amazon (about 90, left blank on purpose: it could be anything), robotics vendors (RevRobotics, AndyMark, Digilent, Seeed), transit (MTA, MetroCard, U-Haul) and plain-English meals ("Friday lunch for hack clubbers", "Tribeca Bagels").
After: 272 suggested (48.6%). Three rules cover those. Because they were written from this same data, the second number is tuned on itself, not held out; the honest figure is the first. Restaurants-by-POS-prefix (TST*, SQ *) and big box stores (Walmart, Target) stay blank, since the text does not say what was bought.
The feed is fetched at run time and not copied. Rerun with node ventures/ai-bookkeeper/eval/hcb-memos.mjs --orgpage 2 --from 0 --to 40.