Annika Brundyn is a solutions architect at OpenAI who has worked on foundation models for language, chemistry and humanoid robotics. Before joining OpenAI, she spent approximately four years at NVIDIA, contributing to Nemotron-4, nach0 and GR00T N1.
Brundyn studied actuarial science and statistics at the University of Cape Town, participated in the 2019 Data Science for Social Good fellowship and helped organize Deep Learning IndabaX South Africa. She subsequently earned a master’s degree in data science at New York University, where she researched three-dimensional reconstruction of surgical video. As first author of a study of neural reconstruction for endoscopic footage, she investigated whether generated stereo views could improve surgeons’ depth perception during minimally invasive procedures.
At NVIDIA, Brundyn worked across embedded systems, computer vision, graph neural networks and privacy-preserving federated learning. Her work on financial-services data silos describes how institutions can collaboratively train fraud-detection models without pooling sensitive customer records; she also developed an accompanying PyTorch and NVIDIA FLARE implementation.
She later coauthored technical reports on the Nemotron-4 15B and Nemotron-4 340B language models, alongside nach0, a multimodal chemistry foundation model integrating scientific text and molecular representations. Her work on GR00T N1 extended foundation-model development into robotics, where she focused on the practical barriers to training and deploying general-purpose humanoids.
- Physical AI for human environments: Healthcare, construction, manufacturing and transportation require systems that act in the physical world. Humanoid designs can operate within buildings and workflows already organized around human bodies.
- The robotics data pyramid: High-quality teleoperated demonstrations are expensive and scarce; internet videos provide scale without reliable robot-action labels; synthetic data can expand useful training examples but requires difficult, carefully constructed simulations. Brundyn identifies combining these sources as a central challenge for generalist humanoid robots.
- Simulation-to-edge deployment: Robotics systems depend on three computationally distinct stages: generating or simulating training data, training foundation models and deploying efficient models directly on robots. Brundyn connects these stages to NVIDIA’s OVX, DGX and AGX infrastructure.
