About the Role
As a Software Engineer on the Machine Learning Infrastructure team, you will build the "Operating System" for our large-scale GPU clusters. You will architect a high-performance training platform that handles the immense complexity of multi-thousand GPU workloads, ensuring every cycle is used efficiently. Your work directly determines the velocity at which our researchers can train and iterate on the world’s most advanced models.
The ideal candidate is a systems expert who thrives on solving the orchestration, networking, and reliability challenges that emerge at massive scale. You will partner closely with researchers to build a seamless, resilient environment that transforms raw compute into breakthrough AI.
You will:
- Architect and scale a multi-tenant orchestration layer that abstracts away the complexity of GPU clusters, ensuring high utilization and seamless job recovery.
- Design and implement scheduling primitives to optimize the lifecycle of training jobs.
- Develop deep observability and automated health-checking into the training stack to proactively identify and isolate hardware failures.
- Evaluate and integrate emerging technologies in the CNCF and AI ecosystem (e.g. Ray, Kueue), making data-driven build vs. buy decisions that balance velocity with long-term maintainability.
- Work closely with Finance and Procurement teams to drive our capacity planning process.
- Participate in our team’s on call process to ensure the availability of our services.
- Own projects end-to-end, from requirements, scoping, design, to implementation, in a highly collaborative and cross-functional environment.
Ideally you'd have:
- 5+ years of experience in backend or infrastructure engineering, with at least 2 years focused on orchestrating ML workloads at scale (100+ GPU nodes).
- Strong programming skills in one or more languages (e.g. Python, Go, Rust, C++).
- Experience with complex compute management systems that cover queueing, quotas, preemption, and gang scheduling.
- Experience with distributed training infrastructure, such as EFA, Infiniband, and topology-aware scheduling.
- Experience with distributed storage systems (e.g. Lustre, S3) as they relate to training throughput.
- Expert-level knowledge of Kubernetes internals (Custom Resources, Operators, Admission Controllers) and how they interact with device plugins for specialized hardware.
- Familiarity with cloud infrastructure (AWS, GCP) and infrastructure as code (e.g., Terraform).
- Proven ability to solve complex problems and work independently in fast-moving environments.
Nice to haves:
- Experience with distributed training techniques such as DeepSpeed, FSDP, etc.
- Experience with the NVIDIA software and hardware stack (CUDA, NCCL).
- Experience with PyTorch.
- Familiarity with post-training algorithms such as GRPO, and with Reinforcement Learning.