Objection 01
“Our line runs at 320 parts per minute - your model needs to make a decision in under 40 ms. Can it actually do that on an edge box?”
Yes, and we will prove it on your line. The TensorRT-distilled model dispatches in under 22 ms on Jetson AGX Orin 64GB at 4K input, leaving margin to write the reject event to the PLC inside the line's takt. Higher throughputs (>400 ppm) move to NVIDIA IGX Orin or a multi-Jetson load-balancer. We publish the latency budget by layer and we measure it on your hardware during commissioning.
Proof: anonymised UAE packaging-plant reference - 12 defect classes, 99.1% detection rate on the historical reject library, <22 ms inference latency at 320 ppm sustained.
