Greykite design question

I wanted to bring up a question regarding the software architecture rather than a specific issue.

While exploring the Greykite source code, I noticed that the final step of the pipeline involves a SilverkiteEstimator object, which wraps SilverkiteForecast. It seems like SilverkiteForecast handles various tasks related to feature transformation and preprocessing, such as Fourier seasonalities, changepoint detection, and time-related feature extraction.

I was wondering if it would be more logical to have each of these steps as separate transformers in the pipeline, instead of consolidating them all in the last step. This approach would not only facilitate the use of Greykite as a feature extractor (by stopping the pipeline before the actual ML step), but also enable the utilization of virtually any sklearn-compatible regressor.

I understand that implementing this change would require significant effort, but I wanted to inquire whether other users share a similar sentiment or if I might be overlooking something.

Once again, I appreciate the great work you've done!

Start by reading greykite/sklearn/estimator/silverkite_estimator.py and greykite/algo/forecast/silverkite/forecast_silverkite.py, focusing on how preprocessing and forecasting steps are combined. Determine whether those steps can be separated into sklearn-compatible transformers and what interfaces would be needed to support feature extraction and alternate regressors; done would require an agreed design and scope.