Manufacturing industries need fast and accurate visual inspection systems that work on low-power edge devices. We present Edge-Optimized Vision, an open-source system designed to detect product defects on affordable AIoT hardware. Using an industrial dataset, our approach follows four key steps: First we convert pixel-level defect masks into bounding box annotations, second we apply realistic factory conditions through data augmentation including blur, low lighting, haze, surface smudging, and partial occlusion, third we train a compact MobileViT-based YOLO detection model, and fourth we compress the model through structured pruning, quantization-aware training, and layer fusion to create an eight-bit deployment-ready version. We introduce EOV-Metrics to measure practical performance, including frame-to-decision time, energy consumption per detection, MQTT communication delays, and accuracy under both clean and degraded conditions. Our system achieves real-time operation on Raspberry Pi hardware. This work combines transformer-based architectures with comprehensive model compression and practical deployment evaluation in a fully reproducible AIoT framework.