• Developed a vision–language pipeline using BLIP-2 and LLaVA to detect accessibility hazards from images, applying spatial reasoning and prompt conditioning to generate clear, context-aware safety explanations for visually impaired users, evaluated through human feedback