Distribute Tensorflow - Training Neural Networks on Client Device

I have a Windows C++ application. It is to be extended with a neural network function that allows end users to train neural networks themselves on their devices (also Windows OS). The neural network modelling can be done automatically and is quite simple.

Tensorflow is well suited to my needs, but I only need a tiny fraction of the functionality. Tensorflow Serving is designed to distribute ready trained networks instead of training new ones.

So far, the only option I see is to distribute pre-built Tensorflow DLLs, which are very large.

How would I achieve my goal? Is there any reference software that has a similar use case? How do they do it?

Thanks a lot!