Why not run a couple PyUNO workers on EC2 and pull conversion jobs out of a MQ system where they get queued up by whatever system does the conversion requests? The workers pull a job out of the queue, load the files from wherever they are (S3 for example) and then put the converted files back when done. With a little scripting the PyUNO workers can be monitored and restarted after <x> conversion jobs, or simply when the conversion doesn't happen within a timeout. And the whole thing can be scaled up simply by running more EC2 conversion workers. This is the sort of thing AWS was made for.
I built this library to deal with OpenOffice's PyUNO service hanging, leaking memory, and not being thread-safe. I think it could be quite useful for people running into similar problems.