VisionGuard: Industrial AI Inspection on AMD

VisionGuard is an industrial AI inspection platform that turns factory camera images into real-time quality decisions: PASS, ALERT_OPERATOR, or STOP_LINE. It uses Qwen2.5-VL-7B-Instruct served through vLLM on AMD MI300X with ROCm, wrapped inside a production-style FastAPI and React/Vite inspection console. The system allows operators to upload product images, inspect sample defects, view annotated outputs, and receive structured defect intelligence such as defect type, severity, likely cause, recommended fix, and factory-owner summary. Every inspection is logged into SQLite, powering event history, shift reports, operations alerts, and runtime metrics. VisionGuard also includes a Factory Adaptation Studio that shows how a factory can onboard its own dataset, analyze readiness, estimate a LoRA/adapter path, register factory-specific models, and route inspections through specialized model profiles. For our real-world factory extension, we added NanoDefects, a dataset of steel-bottle defects from an actual bottle manufacturing line, to evaluate false-PASS risk and build a safer factory QA workflow. The project demonstrates how AMD MI300X can power practical multimodal AI for manufacturing, moving beyond a simple model demo toward a deployable industrial quality intelligence platform.