I've recently come across this visualization of k-means clustering, via a co-worker. It's a good illustration of the influence the starting points have on the final clusterings.
I thought it would be interesting to build something interactive with datasets created on the fly and giving the user the ability to choose k.