We provide a wide range of nearly optimal guarantees for several fundamental problems in robust supervised learning based on a single iterative polynomial filtering algorithm.