Natural and technological disasters continue to threaten communities, infrastructure, and the environment, and the cascading nature of contemporary risks complicates their assessment. This study presents a disaster risk analysis model coupling Geographic Information Systems (GIS), ensemble machine learning, and AI-driven interpretation tools, collectively termed GeoAI, to support multi-hazard risk assessment, resilience planning, and citizen-oriented risk communication. Rather than building isolated models per hazard, the framework applies a single, modular pipeline consistently across flood, wildfire, earthquake, landslide, drought, and urban heat island hazards. Hazard, vulnerability, and exposure layers are derived from open geospatial datasets and processed in QGIS, after which Random Forest and XGBoost classifiers generate hazard and vulnerability maps validated using AUC-ROC, F1-score, and Cohen's Kappa. A distinctive component is an AI agent built on Large Language Models (LLMs) and the Model Context Protocol (MCP), which queries structured risk databases and produces region-specific narrative risk reports in natural language. Outputs are delivered through a web-based platform with an MCP-connected chatbot, lowering the barrier to understanding complex risk information for experts and the public. The model was tested in Türkiye's Marmara Region, which is important due to its dense population, heavy industry, and earthquake risk along the North Anatolian Fault. The pilot showed good results: XGBoost performed better than the baseline models, and the LLM-based interpretation layer gave clear, well-grounded outputs. The framework offers a scalable, open, interoperable approach linking spatial analytics, machine learning, and generative AI for evidence-based disaster risk reduction.