We will describe our work on the fundamental science and implementation of machine learning for both RF sensing and learned physical layers, as well the integration of those technologies into commercial software. Replacing traditional signal processing with machine learning-based solutions reduces power consumption and improves density, throughput, and accuracy. Making these techniques available to many real-world radio systems requires optimized software frameworks and low-SWaP devices. We'll discuss our experience and results in migrating machine learning-based wireless processing products to TensorRT and the NVIDIA Xavier embedded processing device, and show performance comparisons relative to other platforms and frameworks. We will also share some of our work investigating the application of these new technologies to future 5G and IoT systems.