I need a computer-vision model that can automatically flag visible defects or damage on industrial machine images coming from our production floor. All input will be still photos of the equipment; no product shots or assembly-line wide views—just the machines themselves. You will take a dataset that I provide (thousands of labeled examples plus a folder of unlabeled incoming frames) and train a deep neural network to differentiate normal from defective states, then package the result so my team can run inference on a GPU workstation. PyTorch or TensorFlow is fine as long as the final deliverable includes: • The trained model weights • A Python inference script that accepts an image path and returns a binary normal / defect label with confidence • A short README explaining environment setup and how to retrain with new data Accuracy matters more than model size, but I still need sub-second inference on an NVIDIA RTX 3060. I’ll evaluate your work on a hidden test set; anything above 95 % F1 will be considered a pass. If you have experience with industrial inspection, transfer learning on ResNet/EfficientNet, or segmentation-based approaches, let me know in your proposal along with a quick outline of your plan and timeline.