7 days ago I started building a real AI platform: Project 01, a fraud detection evaluation harness on 284,807 transactions with only 0.17% fraud. Accuracy is useless here, so I built an evaluation suite first: precision, recall, threshold analysis, average precision. Then I trained 3 models — Logistic Regression, Random Forest+SMOTE, XGBoost — all measured with the same code. XGBoost cut missed fraud from 52 to 14, catching 38 extra cases and improving detection by 21%.

Check it out