nawnie/Model-Operating-Kernel ? reverse-engineered prompt

Build me a local first Python app that acts like a control layer for AI models on my own machine.

I want it to register different experts, choose which one should answer a request, keep an eye on memory so it can free up unused models when VRAM gets tight, and call either local models or HTTP based backends. It should also save traces for each request, support simple rule based routing now, leave room for smarter routing later, and include a few basic evaluation scripts so I can see whether the routing choices are any good. I’d like a small companion mode too, so I can wake it up, put it to sleep, and run it from the terminal. Please make it runnable with a simple command, include a few sample configs and tests, and look up current docs online if you need to.