First thoughts:
1. How much training does this need to be considered functional? I.e. I see your example GIF of writing two sentences and out comes a formal Cursor email. Looks great, but that goes beyond NLP, there is context in the output I do not see in the input. So what would need to happen to get to the GIF if I start from scratch?
2. Is there any kind of confidence score that indicates the estimated quality of the output? Reason here is that communication is still a pareto system for me: There might be occassions where it's delicate to communicate correctly, and messing up that one time might have disproportionate effects. Granted, I'd reread an email in such a case, but any visual cue as to how likely I should be rereading an output would bring peace of mind.
In any case, congrats on shipping, and best of luck for the next steps!