This is a submission for the Hacktoberfest Challenge, Week 2: Side Hustle.

What I Built

Freelancers collect receipts and invoices from many places: email PDFs, phone photos, app screenshots. At tax time someone has to type each one into a spreadsheet. Those documents hold client names, rates and payment details, and many freelancers don’t want them uploaded to a service they don’t control.

Local Receipt Sorter is a small app that runs entirely on the user’s own computer. You drop in receipts, it reads the vendor, date, invoice number, currency, subtotal, tax and total, flags anything suspicious, and exports a CSV or Excel file for the accountant. It also suggests a clean filename for each document, like 2026-03-04_acme-cloud_1180.00.pdf.

The side hustle idea: sell it to freelancers as [a one-time license / a monthly fee of X / a done-for-you service for X per batch].

Demo

Code

How I Built It
Model: a Gemma vision model (gemma3:4b) served locally through Ollama. Documents never leave the machine.
Structured output: the model’s answer is forced into a JSON schema, so every field comes back in a predictable shape.
Checks instead of self-reported confidence: small models are poor at judging their own accuracy, so I didn’t ask the model how sure it was. A rule-based checker flags missing fields, bad or future dates, non-positive totals, and rows where subtotal plus tax doesn’t equal the total.
Review before export: every row appears in an editable table, so the user fixes mistakes before the file is exported.
Stack: Python, PyMuPDF for PDFs, Streamlit for the interface, pandas for export.
What I Found

I tested it on 25 documents: 15 PDFs, 10 photos and 15 screenshots. .

One result I liked: my first test files were order pages that can generate an invoice, not actual invoices, and they had no date. I assumed the empty date column was a model failure and loosened my prompt. Then I realized the model was right: there was no date to read. I changed the prompt to say “never invent a date, vendor or amount” so it stays that way.

Why Open Innovation Matters
Privacy is the product. The documents are read by a model running on the user’s own computer.

I turned off Wi-Fi and it still worked.
No per-document fee. There’s no API bill that grows with each receipt, so a flat price can work.
I can change it. I can swap in a larger Gemma model with one setting, tighten the prompt, or fine-tune later for a specific kind of invoice.

I did not compare it against a closed model, so I’m not claiming it’s better. The claim is smaller: a small open model, running locally, can do a useful part of this job.

Limits
I tested only 25 documents, so these results show it works, not how often.

What’s Next

More test documents, a larger model, fine-tuning for a niche such as GST invoices, a first paying user

Prize Category

Best Use of Gemma. The whole project runs on Gemma through Ollama.