This project presents an intelligent fire evacuation framework that integrates multimodal data analysis, crowd modeling, and AI-based planning to simulate and optimize real-time evacuation strategies. The architectural blueprints of buildings are preprocessed using deep learning-based semantic segmentation to extract critical spatial entities such as walls, corridors, exits, and hazard zones. These entities are then structured into a spatial graph representation of the environment. To mimic real-world crowd behavior under panic conditions, a social force model is employed, which simulates dynamic movement based on individual demographics, psychological stress, physical constraints, and mobility impairments. Emergency context, including sensor-triggered alerts and crowd reports, is processed using prompt engineering techniques to generate a situational narrative. This narrative is then passed through a Retrieval-Augmented Generation (RAG) pipeline that fetches relevant evacuation knowledge from web-scraped sources and institutional databases. An LLM (Large Language Model) subsequently generates context-aware evacuation policies. These policies, alongside the spatial graph and dynamic crowd data, are input into a Graph Neural Network (GNN) module that predicts optimal evacuation paths and zone-based strategies, accounting for congestion, accessibility, and time efficiency. Simulation results demonstrate a near-optimal path efficiency score of 0.93 and effective hazard avoidance in synthetic environments. However, real-world validation remains a challenge, as the current framework relies on synthetic blueprints and simulated scenarios. Future work includes integration with IoT-based sensors, incorporation of real fire drill data, and deployment within actual infrastructure such as public buildings and transportation hubs to evaluate system performance, generalizability, and scalability in real-world conditions.