Testing a bookkeeping categorizer on 631 charges it had never seen

ai-bookkeeper's keywords suggested an account for 44.8% of 631 real Hack Club card charges, then 50.2% after rules written from the misses.

Tools and projects · 1 min read · Source
venturesai-bookkeeperevaluationbookkeeping

ai-bookkeeper files bank lines with rules, payee memory, a learned model and built-in keywords. To see how the keywords do on merchants nobody wrote a rule for, we read a fresh page of Hack Club's public HCB feed (transparent organisations only; fetched at run time, never copied): 631 card charges.

  • Before any new rule: 44.8% got a suggested account (283 lines).
  • The misses were a long tail: MailerLite, Tello Mobile, SHELL/SHELL, Alro Steel, Bambu Lab, CheapOair and similar. Rules for those take it to 50.2% (317). That second number is tuned on the same data, so it is a ceiling, not a forecast.
  • Bare amazon is 150 of the 631 charges and stays blank on purpose: the text cannot say what was bought.

These memos carry no category, so this measures coverage, not whether the suggestion is right. A labelled set of real books is still the thing we want.

More in Tools and projects

All writing · All in this topic · All tools and research