From 2002 - 2005 I've worked on a modular VJ engine using only Shockwave 3D generated realtime graphics. Every visual was implemented as plugin and received events generated by the audio analysis engine. This video is only making use of one such module (the last one I've ever done). It is based on a Bayesian network to chain audio events in time (taking into account previous events) and alter the probabilities of their action choices to create many different (yet still predictable) outcomes to the same events.
Audio: Interlomas by Damiak (n5MD records)