Visual inspection built for your real-world images.

Accuracy depends on the quality and consistency of your images, the number of labeled training examples, and how well your test conditions represent production. We do not quote accuracy numbers before seeing your data. During the Vision Discovery Sprint, we train a prototype model on a sample of your actual images and report precision, recall, and F1 scores by class — so you have real numbers for your environment before committing to a production build. Clean, well-lit images with sufficient training examples typically reach production-usable accuracy faster than noisy or highly variable image sets.

No. We design the integration layer around your existing camera infrastructure — IP cameras, RTSP streams, industrial cameras, or batch image uploads from your existing capture process. The vision system connects to what you already have. Hardware changes are only recommended when your current cameras produce image quality so low that accurate detection is physically impossible (e.g., resolution too low to distinguish defect types at operating distance), and we will tell you that clearly in the Discovery Sprint rather than after you have committed.

Azure Computer Vision provides pre-trained models for general object detection, OCR, and image tagging — useful for common, well-represented categories. A custom-trained model is trained specifically on your objects, your defect types, your environment, and your labeling schema. For most industrial and domain-specific applications (manufacturing defects, custom product SKUs, specialized documents), the pre-trained model is a starting point for transfer learning, not a final answer. We assess which approach — pre-trained API, fine-tuning, or full custom training — fits your use case during the Discovery Sprint.

There is no universal minimum, but as a starting point: defect detection on a single defect type typically requires 200–500 labeled examples per class; multi-class detection with fine distinctions needs 1,000+ per class. Transfer learning from pre-trained foundations significantly reduces the labeled data requirement. During the Discovery Sprint, we assess your existing labeled data (or help you build a labeling strategy) and give you an honest estimate of what you need before training begins. Do not let an unlabeled image backlog block you from starting — we have worked with teams who begin with far less than they expect.

Yes. We export trained models to ONNX format, which runs on a wide range of edge hardware — industrial PCs, NVIDIA Jetson devices, Raspberry Pi 4/5, and Azure IoT Edge nodes. The feasibility depends on your hardware's compute capacity and your inference latency requirement. Edge deployment eliminates cloud round-trip latency, reduces bandwidth costs, and keeps image data on-premises — all relevant for factory-floor and real-time applications. We test model performance on your target hardware as part of the production build and size the deployment accordingly.

Accuracy drift from environment changes — new product variants, lighting changes, camera repositioning — is expected and planned for, not treated as a failure. The retraining pipeline captures corrections from the human review queue and retrain the model on a scheduled cycle (weekly or monthly depending on the rate of change in your environment). When accuracy drops significantly between retraining cycles, the monitoring system alerts your team and can temporarily lower confidence thresholds to increase human review volume until the model is updated. The system degrades gracefully and recovers automatically.