postech-di-lab/METIS — reverse-engineered prompt

Build me a working ML based decision support system like METIS. I want something that can take in different kinds of data together, learn from them in one unified setup, and return useful recommendations through a simple API, especially next item predictions for a custom dataset with a pretrained model.

Please make it feel like a practical framework, not just a research demo. It should support large and growing data, let the model update over time with incremental learning, and include a privacy friendly federated learning option so local user data does not have to be centralized. The main example can be product recommendation like the jewelry and woodworking use cases in the README.

I also want a clean way to run inference and test it end to end, with sample data prep, model loading, API calls, and a short example showing recommendations being generated. If anything is unclear, look up the current docs or infer the missing glue code from the repo structure so the whole project runs smoothly.