You’ll own the GPU compute that powers our model training, evaluation and serving systems.
This role sits in our Infrastructure team, which works at the lowest layers of our stack, from Kubernetes down to the GPU hardware. You’ll run the clusters, GPUs and schedulers that our research, product and platform teams depend on, and manage the capacity we buy from our cloud providers.
What success looks likeSuccess is measured by how well we use our GPUs: keeping production reliable and improving the performance of our training and serving stack.
What you’ll be doingRunning Kubernetes clusters for GPU training and inference
Improving how GPU work is scheduled and placed
Detecting, draining and repairing unhealthy GPU nodes
Managing capacity with our cloud providers
Debugging failures in GPUs, Linux, storage and networking
By month 3, you’ll be handling a share of cluster incidents on your own and will have shipped one change to scheduling or node health, such as automatically detecting and draining unhealthy GPU nodes. By month 6, you’ll own one part of the compute layer outright: scheduling and placement, node health and repair, or capacity and new regions. By month 12, you’ll be responsible for a globally distributed GPU system that shifts capacity between training and inference as needs change.
Who you’ll work withYour hiring manager is our VP of Engineering. Day to day, you’ll work with the Platform Infrastructure team, the research and product teams who use our compute, and our cloud providers and other vendors.
What we think it takesYou’re likely a good fit if you:
Have built and run GPU workloads on Kubernetes
Could design a GPU cluster for training and online RL from scratch
Debug confidently at the OS layer and below (Linux, GPU drivers, networking and storage)
Write strong Python
It’s a plus if you’ve worked as an SRE or infrastructure engineer, or have used:
Slurm
Kubernetes scheduling and autoscaling tools such as Kueue or Karpenter
Multi-cloud orchestration tools such as SkyPilot
AWS services such as EKS or SageMaker HyperPod
High-performance interconnects and collective communication libraries such as EFA, InfiniBand, NCCL or Open MPI
Terraform
You don’t need ML research experience. If you match most of this but not every point, please apply.
How we hireFour stages, about three hours in total: a 30-minute call with our Talent team, a 60-minute coding interview, a 60-minute system design interview and a closing 30-minute conversation with our VP of Engineering. The coding and system design interviews are modelled on real-world problems you’d face in this role.
PracticalitiesThe role is hybrid, with three days a week in our Paris or London office and occasional travel between the two. We offer a competitive package.
