While this approach did strike me as quite elegant in theory, what I found in practice was that a lot of those interfaces ended up just being boilerplate to give me access to the underlying elements of the data which I would have gotten for free in a more strongly typed language.

In other situations, this approach would work well as long as what I was working on was fresh in my mind. However, when I had to come back to some code and extend its behaviour it would be back to digging through it to either understand the structure of the data... or understand the structure of the fancy abstraction I had built up on top of the data. It just didn't seem to be very time efficient in practice for the things I was trying to build.