ReceiptLens: A Local-First Receipt Desk Prototype

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle

What I Built

ReceiptLens is an early receipt-desk prototype for freelancers and independent workers who spend time turning receipts into bookkeeping records. It has a browser dashboard for uploading JPG, PNG, and WebP images, reviewing recent jobs, and inspecting extracted fields.

The problem is real, but this prototype is not yet a validated product: I have not tested it with a freelancer or small-business owner, and I have no payment or willingness-to-pay results to report.

Demo

The public demo is hosted on Render's free plan: https://receiptlens-free.onrender.com.

The free service hosts the dashboard and API, not Ollama or Gemma. The page reports the model unavailable, and uploads return a fallback marked for review. Uploaded files reach Render, so use synthetic receipts only; do not upload real financial documents. The free service also has ephemeral storage, so saved jobs may disappear after a restart.

Code

Source code: github.com/bklimbad123/open-side-hustle-ai

How I Built It

The dashboard and API use FastAPI, with SQLite for local job records. The intended extraction path sends receipt images to Gemma 3 4B through Ollama. In my Codespace, Gemma answered a text prompt, but image requests failed with unexpected EOF; ReceiptLens then returned its low-confidence fallback. The deployed free service has no model runtime, so its upload path is not a working AI extraction demo.

Why Does Open Innovation Matter?

In a working local setup, an open-weight model can let users keep receipt images on their own machine after downloading the model, avoid per-document hosted inference fees, and switch models as their needs change. These are design benefits, not measured accuracy wins over closed APIs. The local vision path still needs to be made reliable and tested. The public Render deployment is different: uploaded files reach Render, so it is only for synthetic data.

Prize Categories

• Best Use of GitHub Copilot: I used Copilot to build and refine the app. GitHub Actions runs compilation and an import smoke test on pushes and pull requests.
• Best Use of Render: Render hosts the public dashboard and API on its free plan. It does not host the Gemma inference runtime; the demo shows that limitation and should only receive synthetic receipts.
• Best Use of Gemma: Gemma 3 4B is the intended open-weight extraction model and was tested locally. Text prompts worked, but image inference failed with unexpected EOF, so this build does not yet demonstrate successful end-to-end receipt extraction.