While having bazel in Debian would be awesome, did you consider just
implementing another build system for tensorflow and/or looking for some
Funny you should mention that. ;) It has been tried and a very talented developer gave up in frustration  . I have no reason to believe that I would be able to succeed where he failed. On the other hand, we have made some very good progress with Bazel and that would vastly simplify packaging both tensorflow (with all of the dependency benefits to other Debian packages) as well as other software out there that uses Bazel to build.
It seems there used to be a contrib/cmake branch up till 1.15 with build
but it is no longer available/supported according to
So (i) that could be a starting point and/or (ii) maybe 1.15 is fine
enough for now?
I think that's what Mo Zhou was working on in ...
I don't mean to dismiss your idea, it would be great to have tensorflow available on a build system already available in Debian! I'm just thinking that it may be easier to just get Bazel in and build everything the way it's designed rather than reimplementing all of that from scratch.