This document describes a two-stage crawler for efficiently harvesting deep web interfaces using adaptive learning. The first stage uses a smart crawler to locate relevant sites for a given topic by ranking websites and prioritizing highly relevant ones. The second stage explores individual sites by ranking links to uncover searchable forms quickly. An adaptive learning algorithm constructs link rankers by performing online feature selection to automatically prioritize relevant links for efficient in-site crawling. Experimental results show this approach achieves substantially higher harvest rates than existing crawlers.