Indian metropolitan cities face severe traffic congestion, a problem critically exacerbated by recurring exogenous shocks such as monsoon rains and festivals, leading to hazardous delays for emergency vehicles. Existing traffic management systems often fail under these high-variability conditions as they typically ignore external factors and provide single-objective deterministic routes. This paper proposes a multi-stage framework designed to address these limitations by integrating spatio-temporal traffic prediction with causal incident analysis and multi-objective routing. The core of our approach is a hybrid Temporal Fusion Transformer (TFT) and Graph Neural Network (GNN) model that generates uncertainty-aware traffic forecasts by explicitly incorporating weather and event data. We then employ a formal causal inference model to accurately quantify the real-world delay impact of road incidents. Finally, these outputs feed into a risk-aware routing engine that computes Pareto-optimal paths for emergency vehicles, balancing travel time and incident risk. The proposed framework is expected to deliver significant improvements in forecast accuracy and generate safer, faster and healthier emergency routes.