Looks awesome!

May I ask how you calculate the ambient occlusion?
I'm currently using a randomly rotated kernel (normal oriented hemisphere) with 14 samples and a 5 tap filter but the result is still a bit noisy and not that fast...