Engineering Notes From Production Vision Systems
Write-ups from real deployments: stream stability, inference budgets, quantization trade-offs, and the operational scaffolding that keeps vision services alive past the demo.
RTSP Streams That Don't Drop Frames
Most industrial vision projects fail at ingestion, not inference. A field guide to stable multi-camera decoding, bounded queues, and the reconnect logic nobody writes until a night shift burns them.
INT8 Quantization Without Losing the Accuracy You Paid For
TensorRT calibration is not a checkbox. How to choose calibration data from the factory floor, which layers to keep in FP16, and how to prove the quantized model still catches the defect that matters.
Designing Confidence Thresholds Operators Actually Trust
Automation earns trust by knowing when to stop. Practical patterns for setting thresholds from a cost matrix, routing edge cases to the line operator, and capturing every override as training signal.
Observability for Vision Services Running 24/7
Latency percentiles, confidence drift, GPU thermals, and downstream error rates. The dashboard that tells you a camera is failing a week before the plant manager does.