GiTaDi-CrEaTe/Compressing-Convolutional-Neural-Networks-for-Offline-Educational-Deployment ? reverse-engineered prompt

Build me a small Python project that shows how to compress a handwritten digit recognition model for offline use on cheap hardware. I want a simple MNIST classifier, then a pruned version and an INT8 quantized version, with a clear way to compare accuracy, model size, and latency against the baseline. Please include scripts to train or load the models, run evaluations on the full test set, print a quick benchmark summary, and generate a few charts that show the results clearly.

It should be easy to run from the command line on a normal CPU, with tests that confirm the pipeline still works and that quantization does not hurt accuracy too much. If you need to check current TensorFlow or quantization docs online while wiring it up, go ahead. Keep it reproducible and organized so someone can clone it and rerun everything without needing a GPU or internet connection.