A Bank Trained Its Own Foundation Model to Catch Fraud — Freddy Lecue, Wells Fargo

Why watch

Freddy Lecue from Wells Fargo explains how the bank trained a custom transformer foundation model on billions of customer transactions to predict the next transaction. The model tokenizes tabular data (amounts, timing, merchants, zip codes) and can be fine-tuned for fraud detection, credit decisions, anti-money laundering, and marketing. In fraud detection, it outperformed the production XGBoost model with higher catch rates and fewer false positives while maintaining real-time inference speed for card approvals.