chberclaz/MyPT ? reverse-engineered prompt

Build me a fully offline AI platform for sensitive environments, with a web app where people can upload and index documents, chat with the system using those documents, run approved tools, and see a complete audit trail of everything that happened.

I want it to feel operator friendly, not like a toy chatbot. Please include login with roles, a clear workspace for training, inference, retrieval, agent actions, and logs, plus a way to use local models only with no external network calls at runtime. The behavior should be reproducible from saved config and seed values, and the system should favor traceability and control over flashy output.

It should also be easy to run locally with Docker, ideally with GPU support, and I’d like the code organized so the training and web interface are both part of the same project. If you need current docs for any library choices, look them up online first.