going-doer/paper2code ? reverse-engineered prompt

Build me a Python tool that can take a machine learning paper and turn it into a usable code repository.

I want a simple workflow where I can give it either a paper PDF converted to JSON or the LaTeX source, then it should plan the implementation, analyze the paper section by section, and generate a clean project folder with the main code files and artifacts saved along the way. It should work with either OpenAI models or an open source local model, and I want easy scripts I can run from the command line.

Also include a way to evaluate the generated repo against the paper, both with and without a reference implementation, and save the results in a clear output folder. Please make it easy to try on the Transformer example first, and use sensible defaults so I can get it running quickly. If you need to look up current docs or model setup details online, feel free to do that.