Local AI (Gemma 3 4B)Documents never leave the laptopBuilt with Python and Streamlit
Demo video
The problem
Freelancers handle their own money but have no finance team. Receipts and invoices pile up, income is irregular, and uploading client documents to a cloud AI tool is risky.
How it works
1. Upload Drop in receipt images or PDFs.
2. Read A local model extracts amount, date and store name.
3. Check Validation rules flag doubtful records for review.
4. Export Fix the table, then download a clean CSV.
5. Plan See a monthly forecast and unusual-spending alerts.
Measured results
Tested on 7 real receipts of mixed quality (kept private).
6/7amounts correct
4/4dates correct
3/3store names correct
6 of 7flagged for review
~3 minper receipt on CPU
What failed
One amount was read with a wrong minus sign. The validation rule caught it.
An unreadable receipt still got an amount and a wrong store name.
Store names were missed on three receipts.
Limitations
Small test set, and slow on a laptop CPU.
The forecast was tested on synthetic data and used a 3-month average fallback, not TabPFN.
PDF support tested on 1 file. No duplicate check or database yet.
Next steps
Local forecasting, multi-currency support, duplicate detection, late-payer alerts and a bigger test set.
This tool organizes data and gives estimates. It is not tax or accounting advice.