tqchen/tinyflow ? reverse-engineered prompt

Build me a tiny deep learning system in C++ that feels like a small TensorFlow clone, with a clean computational graph API, automatic differentiation, shape and type inference, and a simple training and inference runtime that works on both CPU and GPU.

I want it to be minimal but real, so it can run the included MNIST style examples and show how the graph is built, executed, and optimized. Please keep the code easy to read since this is meant as tutorial code for learning how deep learning systems work from the inside. A small Python front end for running examples would be great too.

If it makes sense, use existing docs or current references online to fill in any missing details, but keep the project lightweight and focused on the core ideas.