What I learned putting TabPFN into a small business tool during Global Hack Week

During Global Hack Week, the author built Cashcast, a receivables tool for freelancers, small agencies, and bookkeepers. The tool estimates the chance of late payment, a likely payment date with an 80% range, and the cash expected each week for the next eight weeks. Cashcast is built using TabPFN, a transformer pre-trained on millions of synthetic tables. The author learned several key lessons during the development process.

1. Version choice is a licensing decision. TabPFN ships with different versions of its model, each with its own license. For a paid product, the author chose the v2 version that uses an Apache 2.0 license with an attribution clause.

2. The CPU limit is real and shaped the design. Training on more than 1000 samples is not allowed by default due to slow performance. The author trained on the 1,000 most recent paid invoices, which took 39 to 45 seconds per run on their laptop CPU.

3. Ask the regressor for quantiles, not a number. By requesting 49 levels of quantiles, the author gained a useful feature that allows the model to be conditioned on invoice age. This leads to a more accurate prediction of when an invoice will be paid.

4. Missing values can go straight in. New clients with no payment history can have missing features, and TabPFN accepts NaN values directly without the need for imputation like scikit-learn baselines.

5. Test it against something boring. The author compared TabPFN with four baselines, including logistic regression and gradient-boosted trees. While the baselines matched TabPFN's accuracy, TabPFN's ranges were narrower and more accurate for clients with steady payment habits.