In a final-year project, I trained six models for fraud detection using two public datasets: PaySim (mobile money) and IEEE-CIS (card transactions). Despite achieving high accuracy, the best-performing model was not deployed due to production constraints. The deployed model uses only three features—transaction amount, one-hot encoded channel, and source dataset—because the datasets lacked common fields.

Handling class imbalance required SMOTE, but a fallback interpolation method was used for deployment. The project highlighted a gap between offline metrics and real-world production decisions. The best model, likely XGBoost, was outperformed in practice by a simpler model that generalized better.

This experience taught me that model selection must consider deployment feasibility, not just performance metrics.