Read the Google paper on their chubby distributed locking system using paxos. They had a hard time getting it right - unless you think you're so much better than them. IIRC Leslie Lamport works (worked?) for them.

Use an existing, well used, well scrutinized implementation of a known algorithm.

I am not sure if he is still at the MV office but he was one of the reasons why MSR kept a presence in MV for a while.

But if you're a researcher/had a CS research background, I don't see why not. By definition research results in novel approaches (ideally better) than the current ones. You have to be willing to challenge the consensus.

PhD students who can prove that their algs work should work on it, like the results of this paper: https://raft.github.io

Keep in mind that this is BFT consensus. Until recently, it was a niche area of research with very little real world usage. Even if an existing algorithm like PBFT met their needs, there are no well tested PBFT libraries.