Not the author, but I know about ontologies generally. Non-overlapping is great, but rarely possible because whatever you're trying to organize might not divide up that way. Next best is eigenvectors. You would find the "pure types" of programer such that each real programer would a linear composition of the pure types. This is really handy for doing statistics. But you might run in to trouble if programmers don't have interaction free types. Beyond that you can use more elaborate interaction models, but those are more of an art then a science. They require a lot of domain expertise to design.
