About
Teaching a robot a new task usually takes hundreds or thousands of demonstrations, plus a round of task-specific training. This team is building the opposite: a foundation model that picks up a new manipulation task from far fewer examples, even when the objects have moved.
It’s an early-stage, stealth robotics startup with a small founding team. The model is transformer-based and trained largely in simulation, on large GPU clusters in the cloud.
This is an early hire to own distributed training end-to-end — scaling from tens of GPUs to potentially thousands, with the freedom to challenge the existing stack and shape long-term infrastructure decisions. A hands-on IC seat, not a management one.
What you'll do
- Own distributed training infrastructure and scale it across large GPU clusters
- Push training from tens of GPUs to hundreds, and eventually potentially thousands
- Improve GPU utilisation, throughput and end-to-end training efficiency
- Optimise networking, storage, checkpointing and data pipelines for large jobs
- Profile and optimise across the full training stack
- Make long-term infrastructure and tooling decisions
- Potentially support cloud inference and deployment as the product develops
What you'll need
- Hands-on experience training large transformer models across multi-node GPU clusters
- Deep distributed-training experience (PyTorch, NCCL, DeepSpeed, Megatron-LM/FSDP)
- Cloud-based GPU workloads, ideally AWS
- A track record improving utilisation, networking, storage, checkpointing and reliability
- Strong enough to lead this area independently and level up the existing team
- A hands-on IC who still wants to write code
- Happy to work on-site five days a week in London (visa sponsorship and relocation available)
Bonus
- Large-scale VLA or robotics model training
- Large-scale computer vision transformer training
- Foundation model or LLM training infrastructure
- Ability to contribute to model architecture or transformer design
Shortlisted candidates will be contacted within 48 hours.
