Thanks for the suggestion — that’s a really useful point. I didn’t do an explicit threshold sweep in this iteration. I focused on stratified validation, training-only undersampling, and comparing tuned models using recall, F1 and ROC-AUC. Your point about calibration after rebalancing is important, and threshold tuning on a validation set would be a strong next step. My background is mainly in full-stack development, so this project pushed me into ML and taught me a lot about imbalance, evaluation, tuning and model validation.