Personal safety applications today mostly respond after an incident occurs, which limits their ability to prevent harm. In this work, we develop a proactive safety-risk prediction system that estimates how dangerous a location may become in the near future. The system combines sequential deep-learning models with boosted decision-tree techniques to understand how local crime risk evolves over time and space. Historical crime records, temporal patterns, nearby points of interest, and environmental context are merged into structured spatio-temporal data grids. The proposed approach uses an LSTM network to learn short-term temporal changes in risk at the grid-cell level, while an XGBoost model evaluates spatial and contextual factors to produce interpretable risk scores. An Isolation Forest module is used alongside these models to detect sudden, unusual conditions that may indicate unsafe situations. The outputs of all three models are merged into a unified risk score that updates continuously and highlights emerging danger zones. When the score crosses certain thresholds, the system can issue early warnings, suggest safer travel routes, or escalate alerts if needed. The system is evaluated on real crime datasets using spatio-temporal cross-validation, and performance is measured using metrics suited for imbalanced data such as AUC, Precision@K, and F1-score. Results demonstrate that the system can provide meaningful early-risk signals while maintaining transparency and privacy-aware processing.