Many deep neural networks for image classification and object detection perform well when presented with data from the same manifold as the training data. But when presented with real-world inputs, such as can occur in autonomous driving, video prediction and medical diagnosis, these same networks can return highly inaccurate confidence projections on out-of-domain (OOD) inputs. To solve this problem, we have developed a system which rejects out-of-domain inputs whether they are far from the data manifold or closer to the manifold. In this presentation we explain how teams from launchpad.ai and Onfido built and trained such a system to improve vision performance in the field. We present the architecture of our approach and details of the key components, including the classifier, the detector and the “introspection” net, and summarize the results we achieved.