Applied Scientist, Amazon Transportation Services

Amazon’s Middle Mile transportation network runs on physical assets and operations that generate an enormous volume of imagery and video. The Network Engineering, Scheduling, Technology (NEST) Science team within the Amazon Transportation Services organization is looking for an Applied Scientist with deep Computer Vision (CV) expertise and the versatility to apply machine learning broadly, to help turn that visual data into automated, operational decisions, spanning asset condition, automated inspection, object and component detection, and other visual understanding problems across the network. In this role, you will develop, train, and productionize computer vision models across a portfolio of high-impact use cases and business challenges, owning the scientific approach from problem formulation through production. You will work closely with other scientists, business owners, and engineering teams to advance modeling approaches, design rigorous experiments to validate them, and launch them to production. This is a hands-on applied science role with a broad scope and direct, measurable business impact. Key job responsibilities -Develop and apply visual perception and representation learning across image and video domains, including recognition, detection, segmentation, and tracking, with deep spatial and temporal reasoning enabled by modern deep learning and foundation-model architectures. -Own large-scale model training, fine-tuning, and learning from heterogeneous or weakly supervised data, including self-supervised and semi-supervised techniques. -Tackle real-world CV challenges at scale: large unlabeled datasets, class imbalance, high visual variability, and inconsistent capture conditions (angle, lighting, occlusion, motion blur). -Partner with product/program, operations, and engineering stakeholders to translate operational problems into well-defined CV objectives with measurable success criteria. -Design rigorous evaluation frameworks with explicit precision/recall tradeoffs and operating-point selection tied to real-world business cost, including the cost asymmetry between false negatives and false positives. -Build and maintain custom training and inference pipelines, and partner with engineering teams to deploy models into production. -Frame ambiguous operational problems from first principles and select the right approach for each, applying CV where it’s the best tool and other machine learning or simpler methods where they are not.