

I work in an AI company. 99% of our tech relies on tried and true standard computer vision solutions instead of machine-learning based. It’s just that unreliable when production use requires pixel precision.
We might throw a gradient descent here or there, but not for any learning ops.
I believe Spotify did this back in the day in order to hide as much of their AB testing from Apple who is essentially a competitor due to iTunes.
Having much of the UI delivered via web also makes it easier to deploy updates as no software update is necessary.