This is pretty much what I'm going for :). At this point in the project we have an einsum that is slow but works for all cases and supports AD. Next steps are splitting up contractions into pairwise contractions, finding optimal orders, evaluating the contractions with the best algorithm (so we might dispatch matrix multiplication to BLAS) and supporting different backends (probably just GPU but if more general tensor types could be supported that'd be great).

That's a good resource. I beleive a lot of these optimizations are in TensorOperations.jl but given the head-start, there's probably still some good ideas on the table in that library which people should be learning from.