[LurkerAbove]

I had an idea about explaining ML predictions without direct access to the data or the model: provide basic stats about your data, use them to generate synthetic data, then train a surrogate to predict your model's predictions on the synthetic data.

Could be handy for model risk management and governance, e.g. if you need a challenger model for SR 11-7 without all the hassle of getting access to the original data, getting the black box model set up, and so on. I wrote it because I remember having to create "throwaway" models to show why I needed a better model; it would have been nice to just make a couple of API calls instead.

SDK: https://github.com/proxyml/proxyml-sdk-python

Schema builder: https://github.com/proxyml/schema-builder

Landing page: https://proxyml.ai/