Defunctionalisation is a beautifully powerful technique. I've found it fits really neatly with the initial algebra pattern of data processing as repeated conver

Defunctionalisation is a beautifully powerful technique. I've found it fits really neatly with the initial algebra pattern of data processing as repeated conversion of a "description" structure into more useful "implementation" structures. In particular, my most recent project at work consisted of a system that takes a description of a computation from the user and successively performs optimisation passes on the description until we get a description of something that could run really efficiently, and then evaluating that efficient description into an actual computation that you really can run. If the successively-more-optimised descriptions are all defunctionalised to the greatest possible extent (in effect, deferring any concrete implementation details for as long as possible), you have much broader scope to optimise during later passes (since all the information "remains visible", not hidden away in function calls). Moreover, this approach is extremely testable: initial algebras are built to be re-interpreted, so you can write reference evaluators for them; while defunctionalisation gives you a way of checking output without needing to run any pesky programs.