The "as-a-service" was meant to be funny, but I have found that it is often difficult to find special-purpose scientific software (with adequate build system) mentioned in papers online, without sneakernet/emailing the authors directly. Hosting analysis-runners on the web would make doing a test much easier. (It's getting better; the ALICE collaboration (of which I'm a member) for instance uses github for the analysis code and dependencies https://www.github.com/alisw, but each scientist doesn't need to upload their fitting & plotting code.)

Writing a REST API in front of their software would help with "service-discovery", usability, and of course reproducibility; but there's the issue of sending back many-GB of data per run, questionable amount of customization (but if you allow uploading an ini config file, that should be ok), no institutional history of doing tasks this way, and the fact that this is not what I'm paid to do.

For more on scientific computation differs from a more typical setup, I highly recommend this StrangeLoop talk by Jim Privars: https://www.youtube.com/watch?v=jvt4v2LTGK0