We're experimenting with the https://ambiq.com/apollo510/. At the moment we're using tflite models. It would be great if we could use tinygrad to train and export the model to C code with helium instructions. It would be interesting to compare performance