Start with a brute force algorithm that tries to find a "witnessing set" that a graph is not (k,b) clusterable. 
Show that the lack of such a witnessing set implies that the graph is (k,2b) clusterable.

Once you have this, then try to make it greedy and randomized with a running time independent of n. Argue that the assumption of being
epsilon-far from clusterable implies that you are likely to run into "good witnesses" when sampling vertices at random.