AI Engineer - Large Scale Pre-Training

The role
As an AI Engineer for Large-Scale Pre-Training, you enter as a key member of the early AI team. You will work directly with our AI leadership and researchers to define the infrastructure we use to train large-scale robot foundation models on petabyte-scale datasets across our GPU fleet. This infrastructure is a critical part of our research and development cycle: it must strike the right balance between stability and velocity, enabling fast research iterations based on reliable results. This is a full-time, permanent position.

Key responsibilities

• Scalable Training Infrastructure: Develop software infrastructure to train multi-billion-parameter robot foundation models efficiently on large video datasets across multi-node GPU clusters.
• Data Pipelines: Build high-throughput, reliable data pipelines that stream, decode and preprocess petabyte-scale video datasets, keeping our GPUs fully utilized and supporting deterministic, resumable training.
• Experiment Management: Build the tools to organize, monitor and optimize large training workloads and model sweeps. Remove bottlenecks in our training pipelines and make sure all our runs maintain the highest possible goodput.

Who we are looking for
We seek people who embrace the empirical nature of AI research and want to accelerate our research velocity with best-in-class tools - so that we can all make breakthroughs faster, together.

• Experience: 3+ years of hands-on experience training multi-billion-parameter models on multi-node GPU clusters, including supervising long-running experiments, in-flight debugging and optimizing goodput and MFU.
• Core Skills: Deep expertise in distributed training and training efficiency at scale. In detail:
  • Distributed Training: Data parallelism and sharding (DDP, FSDP/FSDP2, ZeRO), tensor, pipeline and sequence parallelism, and frameworks such as Megatron-Core, DeepSpeed, TorchTitan or JAX/GSPMD.
  • Training Efficiency: BF16/FP8 mixed precision, activation checkpointing, FlashAttention, and profiling with PyTorch Profiler or Nsight Systems. * Fault Tolerance: Distributed and asynchronous checkpointing, automatic recovery, and diagnosing NCCL and hardware failures.
• Core Stack: Python, PyTorch and/or JAX, CUDA, NCCL, and cluster orchestration with Slurm or Kubernetes.
• A plus: High-throughput data pipelines for large video datasets (WebDataset, Mosaic Streaming, Ray Data, DALI/NVDEC); custom kernels in Triton or CUDA; familiarity with multimodal or robot foundation models (VLAs, video transformers, diffusion/flow-matching policies); experiment tracking and sweep infrastructure (Weights & Biases, MLflow, Hydra).
• Mindset: You balance the agility of a startup with the professional rigor required for a global industrial player. You are a curious builder who thrives in a collaborative, fast-paced research environment, and an engineer with a research mindset who understands the workflows of cutting-edge AI research and supports them with rigorous software engineering.