All three look promising. However, you might be pidgeonholing yourself by trying to go back to the "SRS as generalized learning tool". For most people, it would appear as if that's too abstract. You may be much better off focusing on the most generally appealing use case (The name-face ID one sounds the most promising; I can't imagine any people who weren't already self-motivated autodidacts using the first two). In fact, it might turn out to be much better than than the original Anki-as-service app; it appears to me that many more people view "oh god what's her name I just met her a week ago THINK" as a problem than "Oh, hm, it appears I've forgotten how to say 'praying mantis' in Japanese".

To extend the Name-Face identification concept, you could also add things such as people's birthdays, dates of important events such as anniversaries into it; although I'm not sure how many of those things aren't problems anymore because of calenders etc.

I can't comment too much on the "Twilio of online learning" idea; I don't know the interest level of online courses such as Udacity, Coursera and Codeacademy on something like that. Although I will warn that there's a real risk that it'll be treated as "just another complicated feature that I don't need to use" by the average student. But if you get a hardcore userbase who are happy with the product and willing to give feedback then you're in much better shape than trying to arbitrarily design for the "average" user.