If you have experience with deploying machine learning on a large scale, then you already know that the systems that employ machine learning as a basis for generating data are already extensively coupled with the technologies they run on. I guess that would seriously impede with your fun factor if you'd disagree with the overall mix they have in place.
There's always lots of crazy ideas you can go on about with the available domain data with creative research or exploration. Perhaps that's your thing? I would believe that what of it get's into production in big companies does so after proving it's use and getting meticulously scrutinized with resource consumption measurements and estimates by several different people, to the point that even something that might be very exciting for the users might not make it in (perhaps they've calculated that the size of the data could explode beyond what they'd like to invest into). I guess that would seriously impeded with your fun factor if you'd have a hard time associating the obvious potentials and their willingness to go through with it.
I'd just warn you that we don't all get it as good as Mikio Hirabayashi, to both develop the requirements and have complete freedom about the technologies that the system will run on.