Have you considered adding support for human-in-the-loop optimization? I believe this tool could find a lot of use in real-world experimentation.