Shoebox, a new receipt sorter for gig workers and freelancers, is running entirely on users' local machines, eliminating the need to upload sensitive financial data to third-party servers. Submitted by developer johnx4321 for the Hacktoberfest Open-Source AI Challenge, the tool leverages the Qwen3-VL 8B vision-language model via Ollama to automatically categorize receipts against IRS Schedule C line items. This approach targets the tens of millions of Americans who earn independent income, offering a private alternative to cloud-based expense apps that often require subscription fees or compromise user privacy.

Local Processing Architecture

The core pipeline scans a local folder for images, hashing each file to prevent duplicate processing. Each image is resized to 1280 pixels and sent to a local Ollama instance, which enforces a strict JSON schema to ensure the model outputs only valid Schedule C categories. Results are cached in a shoebox.json file within the directory, allowing users to interrupt and resume processing without losing progress. The interface is a simple HTML review page served on localhost:8765, where users can correct misclassifications and export a CSV for their accountant.

Handling Edge Cases and Model Quirks

Development involved navigating specific behaviors of the Qwen3-VL model. The creator discovered that the model would sometimes commit to a category before generating its reasoning, leading to contradictions. This was fixed by reordering the JSON schema to require the 'reason' field before the 'category' field, forcing the model to explain its logic first. Additionally, compatibility issues with Ollama versions 0.34 and 0.40 regarding the 'think' parameter were resolved by increasing the token limit to 3000 and parsing whichever field contained the output, ensuring robust performance across updates.

Privacy and Unit Economics

Shoebox operates on a $19 one-time pricing model, made possible by the zero marginal cost of local inference. Unlike cloud APIs that charge per image, Shoebox incurs no additional costs after the initial model download. This economic structure allows the tool to remain affordable for gig workers who might otherwise pay $150 to $300 for bookkeeping services. The privacy benefit is significant: receipts often contain sensitive personal information, such as pharmacy purchases or dining locations, which remain on the user's hardware.

Key Takeaways

  • Shoebox uses Qwen3-VL 8B via Ollama for fully local receipt processing.
  • The tool maps receipts to IRS Schedule C lines with a one-time $19 price.
  • Schema ordering and token limit adjustments were critical for accurate model output.
  • No internet connection is required after the initial model download.

The Bottom Line

Shoebox proves that local LLMs can solve real-world privacy and cost problems in dev tools, offering a compelling alternative to cloud-dependent SaaS for personal finance.