SafeSight

SafeSight AI is an end-to-end video safety inspection platform built for construction sites, warehouses, and industrial environments. It replaces slow, error-prone manual video review with an automated AI pipeline that detects PPE violations, generates structured evidence, and answers natural-language questions about what was found. Upload any .mp4 worksite recording and SafeSight automatically extracts frames across the video timeline, runs a custom-trained YOLO model to identify safety violations — missing helmets, missing safety vests, unprotected workers — and stores every detection as a timestamped event with confidence scores, risk levels, and AI-generated explanations and recommendations. The inspection results surface in a clean web dashboard where safety teams can browse violations by event, view the exact evidence frame for each detection, read an AI-compiled safety report, and ask follow-up questions in plain English — "Was anyone working without a helmet?", "Should work be halted?", "What were the highest-risk moments?" — answered by Qwen using the inspection evidence as context. SafeSight is built on FastAPI and Ultralytics YOLO on the backend, Next.js and Tailwind CSS on the frontend, and deployed on an AMD GPU droplet via AMD Developer Cloud with ROCm powering the AI workloads. The Qwen model runs through an OpenAI-compatible endpoint on the same droplet, behind an Nginx reverse proxy — making the entire stack self-contained on a single AMD MI300X instance.