nubank/fklearn ? reverse-engineered prompt

Build me a Python machine learning library that makes training and validation feel clean, repeatable, and easy to reason about.

I want the core idea to be functional, so models are created with simple learner functions, then produce predictions in a consistent way, and it should be easy to chain training, validation, and analysis without a lot of glue code. Please include a few common models and make it possible to plug in LightGBM, XGBoost, and CatBoost if those extras are installed. It should also support solid experiment tracking style outputs, so I can inspect results later and reproduce them.

Set it up as a proper package with tests, docs, and examples in notebooks. Make sure the code is easy to extend, and follow a style that would work well for real production use, not just demos. If you need current library docs or best practices, look them up online first.