Email me to unlock the full case study and demo.
CNN → ONNX INT8 on Jetson · 30+ FPS · −18% scrap rate.
We started with a compact CNN, pruned and quantized to INT8, then exported to ONNX for Jetson deployment. We batched frames, tuned memory copies, and used CUDA streams to overlap preprocessing and inference. The result: stable 30+ FPS while maintaining high recall.
A small rules layer post-processes detections to cut false alarms. Metrics are pushed back to the central dashboard so we can watch FPS, precision/recall, and scrap rate in near real-time. A/B rollouts let us ship improvements without line downtime.