Abhilash Bora

Finance Copilot

A local-first personal finance app that parses bank statement PDFs and categorises every transaction on your own machine.

One command to launch. Runs locally, with demo data one click away. Needs uv.

The Finance Copilot dashboard, showing a month of demo data: money in, out, invested and kept, a balance chart and spending by category

Finance Copilot is published on PyPI and runs with one command. It installs, starts on localhost and opens in the browser, with no account and no cloud: the data lives in a single SQLite file on your machine. There is demo data built in, so you can try it without a statement of your own.

It parses bank statement PDFs, checks every row against the statement's running balance and categorises expenses through a multi-stage pipeline.

The pipeline prioritises deterministic rules and previously reviewed transactions before falling back to an LLM when necessary. Predictions are assigned confidence scores, with uncertain results sent for manual review. It works without any model at all, and a local or hosted one can be added when wanted.

  • Statement parsing: a dedicated Kotak parser plus a generic parser for tabular statement PDFs from other banks. Every extracted row is validated against the running balance, so a misread amount is caught rather than stored.
  • Multi-stage categorisation: your own rules, built-in rules and learned merchants first, then retrieval of previously reviewed transactions, then LLM-based classification. The progressively more expensive stages are used only when needed.
  • Confidence calibration: tracks the reliability of each categorisation stage and adjusts confidence based on historical corrections and confirmations.
  • Human review: routes low-confidence predictions to a keyboard-driven review queue rather than silently accepting them.
  • Local storage: uses SQLite and sqlite-vec for transaction data and similarity search. Supports local model inference through Ollama or a built-in llama.cpp model, and OpenAI-compatible endpoints where configured.
  • Distribution: packaged with the React frontend bundled in, so one command gives a working app on macOS, Windows and Linux.

Python, FastAPI, React, SQLite, sqlite-vec, Ollama, RAG

SourcePyPIAll projects

abhilashbora3107@gmail.com