Insurance fraud exerts a considerable economic strain on both insurers and policyholders, eroding the foundational trust that sustains the insurance industry. This study presents the development of a robust, scalable machine learning framework tailored for the detection of fraudulent claims. To mitigate the prevalent issue of class imbalance in fraud detection, a synthetically balanced dataset was constructed using the Synthetic Minority Oversampling Technique (SMOTE). The methodological approach incorporates a Random Forest classifier, augmented with domain-informed feature engineering—such as income-to- liability metrics and other fiscal indicators—designed to enhance discriminative power. Model performance is rigorously assessed through a suite of evaluation criteria, including accuracy, precision, recall, F1-score, confusion matrix, and both ROC-AUC and precision-recall curves. The experimental outcomes reveal exceptional classification efficacy, with the model achieving near-flawless predictive accuracy on the balanced dataset. When benchmarked against baseline models, the proposed system demonstrates superior detection capability, particularly in minimizing false negatives. This research underscores the effectiveness of ensemble-based algorithms in the fraud analytics domain and sets the stage for practical deployment in operational insurance environments. Prospective advancements may involve the integration of deep neural networks and real-time anomaly detection using streaming architectures to further enhance scalability and responsiveness.