So i think the recognition algorithm needs a little work. I know you mentioned font limitations, so i suspect adding more of those would be a good start.
Also, i don't see Japanese support, which is unfortunate. If you could get that working i would be an instant fan. (I deal with transcription of Japanese a lot, and although i'm proficient in kana, my kanji skills are pretty poor and probably always will be. What i've always wanted is a tool that can return kanji characters based on a drawing WITHOUT requiring strict adherence to stroke order. That last requirement rules out every single tool i've found...)
Anyway good luck!
This reminds me of a very useful project, Detexify: http://detexify.kirelabs.org/classify.html
Detexify's popularity and utility shows that this sort of thing could be very useful--I suspect I might end up using it myself in the future.
The ajax urls are borked.
Someone done goofed with the ajax.