School safety remains a critical concern within urban education systems, directly influencing student well-being and institutional performance. In this project, we explored how machine learning models could be used to predict school safety ratings across Chicago Public Schools (CPS) systems, providing school leaders with a data-driven tool for early intervention and resource planning. Using a dataset that included safety metrics, attendance records, disciplinary incidents, and climate survey results from 478 elementary and high schools, we tested three predictive approaches: linear regression as a baseline, random forest for ensemble learning, and XGBoost for gradient-boosted performance. Among these, the XGBoost model performed the best, achieving a tested R-squared value of 0.923, a root mean square error (RMSE) of 5.16, and a mean absolute error (MAE) of 4.08 on the test set. Notably, the model identified family involvement scores, student attendance rates, and reported misconduct incidents as the influential predictors of perceived school safety ratings. These results align with existing research in education policy while providing new insights into measurable factors influencing school safety. The practical implication of this work is a machine learning tool that can identify potentially at-risk schools with 92% accuracy. Importantly, because the model is interpretable, administrators can understand how each factor contributes to the predictive safety score, allowing for more targeted and transparent decision-making. This could support efforts in designing preventative strategies, resource allocation, support staff deployment, or tailoring school-based intervention strategies. Overall, this study illustrates how machine learning can complement traditional methods in school safety evaluation, offering a scalable and evidence-based approach that may be particularly valuable in data-rich but resource-constrained educational environments.