There was a time (6 years ago) where I was interested in reimplementing Sklearn but unfortunately I'm too busy with other projects.
However, I would go like this. 1. Create a robust dataframe library and implement the time-series centric features (like rolling sums and things like that). 2. Create the key fit/regression model. This is something that Nvidia recently had to do with rapids.ai and CuDF. And they did that within the past 3 years so there are a lot of lessons learned there and also strategies to minimize time-to-market. The main issue in porting Sklearn is getting lost in the wild wild west of contributions with varying degrees of maturity.
