Pantograph is training general models that start by watching internet-scale video and end up on robots. We think the path to capable robots runs through general intelligence rather than narrow, robot-specific skills. We're scaling simple methods across video games, real-world video, and our own fleet of affordable, durable robots.
We're looking for a research engineer to help us train increasingly capable models across enormous and diverse datasets.
You'll work across the boundary between research and engineering: implementing new ideas, scaling experiments across large GPU clusters, building the systems that let us iterate quickly, and figuring out why things aren't working. The work spans large-scale model training, multimodal representation learning, reinforcement learning, data processing, evaluation, and the infrastructure required to support all of it.
You might be a good fit if you:
Have trained models across large GPU clusters and are comfortable working with Kubernetes
Have built or operated complex distributed systems
Have worked with multi-terabyte or multi-petabyte datasets
Are comfortable with large-scale data processing tools
Care deeply about observability and collect enough metrics to understand what every part of a system is doing
Are comfortable moving between research code and production-quality systems
Like running experiments, getting surprising results, and digging in until you understand why
Move quickly and reach for simple approaches before complicated ones
Nice to have:
Experience with JAX
Experience writing CUDA kernels or otherwise optimizing GPU workloads
Low-level Linux or kernel programming experience
Experience with large-scale video or multimodal datasets
Experience building training or evaluation infrastructure
Experience with distributed training
Experience deploying models into real-world systems, especially robotics
We care much more about what you've built than any specific credential. We're a small, fast-moving team working together in person in San Francisco. If you're excited about architecting novel systems at unprecedented scale, we'd love to talk.