Built using Python, Scikit-learn, and PyOD, trained on datasets such as the Kaggle Credit Card Fraud dataset. Uses preprocessing, feature engineering (frequency, amount deviation, location mismatch), and algorithms like Random Forest, Gradient Boosting, Isolation Forest, or One-Class SVM. Evaluated with metrics like ROC-AUC, precision, and recall to address dataset imbalance.
A Model Service acts as a wrapper, ensuring scalability and independent updates. Fraudulent transactions trigger alerts published back to Kafka (fraud_alerts) and push notifications (e.g., Slack, Jira, or email).
A Model Service acts as a wrapper, ensuring scalability and independent updates. Fraudulent transactions trigger alerts published back to Kafka (fraud_alerts) and push notifications (e.g., Slack, Jira, or email).
Executive dashboards visualize ROI and fraud trends, with fraud scores, thresholds, and real-time blocking capabilities. The system demonstrates <500ms latency to prove “real-time” capability. Designed as a trusted, business-critical AI solution that combines compliance, risk prevention, and operational efficiency.
