The implementation of TClass was inherently experimental. There are significant optimisations possible. For example, feature extraction need only be done once and can be stored for reuse. It could be reimplemented in a faster language (eg. C++) and take advantage of certain data redundancies and repeated calculations. In particular, many data structures are non-contiguous in memory in the current implementation. By reorganising the data structures in memory to lead to greater contiguity and hence better memory locality it is likely that performance will improve significantly.