This is exactly what we had in mind. The latency of hitting the server for each kanji conversion would make it difficult to compete with a native implementation, but this is just a hack to fill in the holes when no such implementation exists.

Aside from other language support, we're thinking of eventually implementing a passive feedback mechanism that would tell the back end about conversion preferences, so that users could teach the system merely by correcting suggestions.