SafeFlow is an early-warning framework for freeway congestion using Caltrans PeMS sensor data on I-5. At each 5-minute forecast interval, it estimates whether a currently uncongested station will fall below 40 mph within the next 30 minutes. We engineered 29 features describing current speed, flow and occupancy, recent traffic changes, neighboring-station conditions, roadway context, and historical traffic patterns. We compare Random Forest, boosted-tree, and Transformer models using chronological walk-forward validation and a frozen holdout period. Beyond ranking accuracy, SafeFlow evaluates whether warnings arrive at least 15 minutes early while keeping false-alert burden manageable. Boosted trees offer the best combination of ranking performance and computational efficiency for our current features and data, while the Transformer achieves slightly higher early detection. Historically unusual congestion remains much harder to predict.