[lspears]

The more interesting problem is learning doc -> image directly. With Dropbox's scale of data, seems feasible for some data types.

[caturopath]

I'm sure that couldn't leak anything important.