Test household transaction categorization in a local AI spreadsheet
At a glance
A Show HN tester ran Probly locally to categorize bank transactions and found that the generated labels needed substantial correction. This workflow is listed under Personal Finance Tools, Spreadsheets, Show HN.
- Agents / tools
- Probly, Docker Compose, OpenAI API, Pyodide
SourceBuilding a tool for these workflows? See weekly rates.
What they did
1) Have Git, Docker with Compose, an OpenAI API key and transaction data ready. The linked project documents cloning https://github.com/PragmaticMachineLearning/probly.git, entering the probly directory and creating a .env file containing OPENAI_API_KEY. 2) Run docker compose build and docker compose up -d, then open http://localhost:3000. These setup details come from the linked repository. 3) Import the transaction data, open the chat using the speech-bubble control and request household expense categories. 4) Review the labels against the underlying transactions before exporting the sheet. The tester's actual request was to categorize bank transactions; many rows became Other and bank charges were mislabeled Fuel, so the output was not ready to serve as an unchecked budget.
How it went
The tester reports successful local setup, import/export and chat access, but poor categorization with their o1-backed setup.
Worth knowing
Python runs in-browser through Pyodide; the repository also documents OpenAI API calls through a server. Local hosting does not make every processing step offline.