A text message becomes a categorized Notion row

Logging a purchase by hand means opening a separate app or spreadsheet, then picking a category, a currency and who it was for from a set of menus, for every coffee and every bill. That friction is exactly why a spending log tends to fall behind or stop after a few weeks. The bot logs a transaction from wherever it already gets mentioned: a short message in English or Russian.

A message like 'spent $45 on lunch' or 'потратил 500 на обед' is the only typing involved. The bot replies with buttons for who it is for, then a currency choice between USD and EUR, then a preview of everything it parsed: amount, type, category, date and notes. Behind that reply, tools/parse_transaction.py matches the message against 22 categories by keyword, checked in order until one fits, and decides expense or income from words like 'spent' and 'paid' against 'received' and 'salary'.

Category detection is rule-based rather than a model call, so a message that matches no keyword lands in Other instead of a smart guess, and currency is limited to USD and EUR. The bot only answers Telegram user IDs on an explicit whitelist, and re-validates the person and currency server-side even though they arrive from button taps. The same code runs two ways: main.py for a local process, or modal_app.py on Modal, where credentials live as Modal secrets instead of a local .env file.