The increasing complexity and vulnerability of modern supply chains, exacerbated by geopolitical tensions, climate variability, and fraudulent activities, highlights the need for robust AI-driven risk management solutions. This research presents a unified, data-driven framework that utilizes machine learning (ML), deep learning (DL), and reinforcement learning (RL) to enhance supplier risk resilience and optimize logistics under disruptive conditions. We use a comprehensive dataset of 1,000 supplier transactions, enriched with historical demand, weather indices, geopolitical risk scores, shipment anomalies, and financial health indicators. We apply various regression models, including Linear Regression, Random Forest Regressor, XGBoost Regressor, and Multi-Layer Perceptron, to forecast future demand and quantify supplier risk, assessing performance with metrics such as Mean Absolute Error (MAE), Mean Squared Error (MSE), and R². Next, we employ Isolation Forests for real-time disruption detection, analysing features like price spike percentages, delivery delays, and sentiment scores to enable the early identification of anomalous events. To optimize dynamic routing in the face of stochastic disruptions, we design a custom Open-AI Gym environment and train a Deep Q-Network (DQN) agent that balances fuel costs, delays, and penalties for anomalies, evaluating the strategy's effectiveness through cumulative reward analyses. Finally, we built a deep neural network using a synthetic fraud dataset for transactional fraud detection, applying SMOTE for class balancing. This results in near-perfect accuracy (>99.9%), as validated by train/validation loss curves and classification reports.  The integrated framework provides end-to-end supplier risk analytics, combining predictive forecasting, anomaly detection, route optimization, and fraud identification to support resilient decision-making in supply chain operations. Key evaluation metrics include MAE, MSE, and R² for forecasting; contamination rates for anomaly detection; cumulative rewards for reinforcement learning performance; and accuracy, precision, recall, and AUC for fraud classification