It turns out it's useful!
But having all my data sent to an AI lab in realtime feels wrong.
So I built an open-source version instead. Interestingly, local inference worked in my testing though it's certainly lower quality. Once we capture more data, we can explore LoRA adapters to improve local quality even more.
Open to feedback!