Whenever you're ready, walk me through how you'd approach this.
How would you optimize a perception model for deployment on an embedded platform?
Discuss a practical workflow for selecting and modifying the model, compressing it, compiling it with NVIDIA TensorRT, and validating accuracy after optimization. Address latency, memory, power, numerical precision, unsupported operators, calibration data, and differences between desktop and embedded performance. Include the experiments, profiling data, and deployment artifacts you would produce.