FinSense
- Question
- Can I stop typing receipts into spreadsheets and just ask questions about my money?
- In short
- Personal finance you can talk to
- Started
- 2025
- Apparatus
- PythonFastAPIReactPostgreSQLTesseractDeepSeek tool calling
- Specimens
- Source ↗Demo video ↗
- Write-up
- Read the full story of how it works →
Shelved. FinSense shipped as a week-long spike and taught me a lot about OCR, tool calling and Postgres RLS; I’m not running it as a live product anymore. The lab notebook write-up stays.
Curious about
Expense tracking fails because entering the data is boring. What if a photo of a receipt were the whole input, and “how much did I spend on food in March?” were the whole report?
Built
A web app that OCRs receipts (Tesseract, German plus English) and has an LLM turn the text into a merchant, total, date and categorized line items. You confirm the result before it becomes a transaction. A chat interface answers questions about your money through tool calling: the model picks from twelve tools and Python runs the scoped, parameterized queries. It has budgets, categories, charts and a dark mode. The backend is FastAPI on PostgreSQL, and the front end is React, TypeScript and Vite.
Learned
- OCR gets you 80% of the way there. The last 20% is prompts, JSON repair and a human confirmation step.
- For anything involving money, tool calling beats letting a model write SQL.
- Row-level security doesn’t apply to the table owner unless you
FORCEit. A safety feature you haven’t tested is a feature you don’t have. - A two-minute demo video explains more than any README.
next specimen → EXP-012 · AranetBar