Timely graduation is a critical goal for higher education institutions, with implications for student success and institutional efficiency. This paper presents a structured pipeline for preparing and evaluating an institutional dataset built from academic, demographic, and administrative records of Construction Engineering program students at the Pontificia Universidad Católica de Valparaíso, Chile. The dataset was anonymized and preprocessed to support machine learning experiments without class balancing or hyperparameter tuning. Using PyCaret, we benchmarked 15 classification models to assess baseline predictive performance. Results show promising accuracy but limited Recall for the minority class—students who graduate on time—highlighting the impact of class imbalance. We discuss constraints such as excluding non-academic factors and propose future work integrating contextual and behavioral data to enhance predictive power and institutional applicability.