Foundation Modelling Engineer at Uncovr

You will build the foundation model layer that makes manual labeling optional for surgical video. By training self-supervised video encoders and spatiotemporal transformers on raw 2-6 hour surgical footage, you will enable machines to understand tool-tissue interactions and procedure progression, directly shaping the future of autonomous surgical robotics and AI-driven operative reporting.

• Train self-supervised video encoders using masked video modeling and temporal contrastive learning on laparoscopic and robotic video datasets.
• Build spatiotemporal transformers to track tool-tissue interactions and construct scene graphs that describe surgical actions independent of tool brands.
• Develop production-grade training infrastructure and data pipelines to turn massive hospital video streams into scalable, high-performance training workflows.

• 5-8 years of hands-on experience training large-scale vision or video models, with deep expertise in representation learning and self-supervised training.
• Proficiency in optimizing distributed PyTorch training for long-context temporal modeling and handling multi-hour video sequences.
• A proven builder mindset with the ability to bridge high-level research and production-grade engineering, ideally supported by publications at CVPR, NeurIPS, or ICLR.

• Work with a world-class team from ETH Zurich, ESA, and Oxford Dynamics, advised by Prof. Dan Stoyanov, a global leader in surgical computer vision.
• Tackle one of AI's hardest frontiers: training large-scale foundation models on complex, long-context video data where labels are traditionally scarce and brittle.
• Join an early-stage startup as one of the first ten employees, gaining meaningful equity and the autonomy to design technical architectures from scratch.