Build the foundation.
We pre-train physical intelligence models for sovereign nations. We’re building a small, senior team with the depth to take a model from first principles to a reproducible checkpoint.
Three core roles
Principal Research ScientistPhysical model pre-training
The work
Own the pre-training research agenda: temporal representations, predictive objectives, and transfer across embodiments. Turn questions about world-action and vision-language-action models into controlled experiments and reproducible training recipes.
The bar
- 5+ years of substantive ML research, with a PhD or an equivalent record of original research.
- Led a large-scale multimodal, video, or robot-model training effort, including ablations and failure analysis.
- Depth in optimization, sequence modeling, and experimental design; strong implementation ability in PyTorch or JAX.
Evidence of ability
Research you led, code you wrote, and a clear account of which experiments changed your conclusions.
Staff ML Systems EngineerDistributed pre-training
The work
Own the training runtime: multi-node execution, memory and communication efficiency, checkpoint recovery, and reproducibility. Build a stack that a national team can operate and continue developing independently.
The bar
- 7+ years of systems or ML engineering, including ownership of sustained training runs across multiple GPU nodes.
- Deep experience with sharding, tensor or pipeline parallelism, mixed precision, and collective communication.
- Ability to diagnose bottlenecks from GPU traces, reason about CUDA kernels and NCCL, and resolve training instability.
Evidence of ability
A system you owned, measured throughput or reliability improvements, and a difficult training failure you diagnosed.
Staff Research EngineerTraining data & evaluation
The work
Own the connection between data and model capability. Build reproducible video and robot-trajectory pipelines, improve training mixtures, and design evaluations that expose weak transfer before a checkpoint is handed over.
The bar
- 5+ years of ML research engineering, with ownership of datasets and experiments used in large-model training.
- Depth in temporal alignment, action representations, deduplication, provenance, and efficient data loading.
- Strong Python and ML engineering; rigorous evaluation design, leakage prevention, and data-mixture ablations.
Evidence of ability
A dataset or evaluation system you built, and measured evidence that a data decision improved generalization.
Application details will be posted here.
Explore our research direction →