Freelancers often drown not from a lack of work, but from the gap between delivering it and getting paid. Paydar, a new open-source project submitted to the Hacktoberfest Open-Source AI Challenge, aims to close that gap by running entirely on the user's laptop. The tool uses local AI models to read invoices and predict cash flow, ensuring sensitive financial data never leaves the machine.

Local Inference for Privacy and Cost

Paydar leverages Gemma 4 E2B, a 2-billion-parameter model running locally via Ollama, to extract data from PDFs, emails, and photos. The developer implemented constrained decoding using JSON Schema to ensure reliable data extraction without malformed dates or missing fields. By turning off the model's "thinking" mode for simple extraction tasks, the tool reduces CPU overhead, making it viable on standard 8GB laptops.

Predicting Late Payments with TabPFN

The system uses TabPFN-3.5, a tabular foundation model, to analyze payment history and predict which invoices will be late. Unlike traditional models that require extensive training, TabPFN works in a single forward pass, making it ideal for the small datasets typical of solo freelancers. It generates 17 features per invoice, including client track records and month-end timing, to forecast both the probability of a late payment and a likely pay-date range.

Practical Workflow and Pricing

The pipeline is orchestrated using Mastra, a TypeScript workflow framework, which breaks the process into five typed steps: scanning, predicting, forecasting, drafting reminders, and reporting. The tool drafts reminder emails timed to risk, sending friendly heads-ups before due dates for clients likely to pay late. Because the heavy lifting occurs on local hardware, Paydar is priced as a small utility at $6/month or a one-time $39, avoiding the high costs of cloud-based SaaS seats.

Key Takeaways

  • Paydar runs Gemma 4 E2B locally via Ollama to keep invoice data private and reduce API costs.
  • TabPFN-3.5 predicts late payments using a single forward pass, suited for small freelance datasets.
  • The tool is built with TypeScript and Mastra, offering a $6/month or $39 one-time pricing model.
  • Sensitive data stays on the device, with only pseudonymized numbers sent to the TabPFN API if used.

The Bottom Line

Paydar proves that local-first AI can solve real business problems without compromising privacy or breaking the bank. For freelancers living on the edge of cash flow, this tool offers the visibility needed to avoid debt without relying on expensive cloud services.