Predicting student academic success allows for timely interventions to improve outcomes. In this work, we develop a hybrid ensemble model that combines a Random Forest (RF) and a Gradient Boosting (GB) classifier for student pass/fail prediction using the UCI Student Performance datasets. We first merge two datasets (Mathematics and Portuguese language) of secondary school students, yielding 1,044 records, and define a binary target: pass if final grade ≥ 10 (out of 20). After thorough preprocessing (duplicate removal, encoding, etc.), we apply mutual information to select the top 10 features most predictive of success (notably G2, G1, prior failures, etc.). The RF–GB ensemble is constructed by calibrating each model’s probabilities via Platt scaling and then learning an optimal weighted average (via logistic regression) of their outputs. We compare the hybrid model against standalone RF and GB on a hold-out test set using Accuracy, Precision, Recall, F1, and AUROC. The results show that the hybrid ensemble achieves performance on par with the strong RF baseline (e.g. ~90% accuracy, F1 ≈0.93), and modestly improves calibrated probability estimates. We discuss the interpretability of the model (feature importances and SHAP explanations) and practical deployment considerations. Overall, this work demonstrates a lightweight yet effective ensemble approach for student performance prediction, with potential to inform early warning systems in education.