We are seeking a Staff AI Orchestration Engineer to lead the design, optimization, and scaling of our Kubernetes-based AI infrastructure. In this role, you will tackle the unique challenges of massive-scale AI workloads, focusing on throughput, GPU utilization, and fault tolerance to support next-generation distributed training and disaggregated inference.
What You'll Do:- Architect Large-Scale Scheduling: Design and optimize hierarchical, high-throughput scheduling architectures for massive Kubernetes clusters (1,000+ nodes, 10,000+ pods), utilizing techniques like optimistic concurrency, multi-scheduler architectures, and batch dispatching.
- Maximize GPU Utilization: Eliminate GPU waste in multi-tenant environments by implementing fractional GPU allocation, leveraging mechanisms like KAI-Scheduler's Reservation Pods or hard-isolation tools like HAMi, and configuring time-based fairshare scheduling to balance over-quota pool access.
- Optimize Placement & Topology: Deploy topology-aware scheduling to align pod placement with physical hardware dimensions, such as NVLink connections, PCIe lanes, and NUMA nodes, minimizing communication latency for multi-GPU operations.
- Enhance Cluster Performance: Reduce scheduling latency and API server load by tuning etcd, optimizing admission webhooks, and implementing in-place pod resizing (VPA) or in-place container restarts.
- Secure AI Workloads: Design secure, multi-layered isolation environments and Agent Sandboxes to safely execute untrusted LLM-generated code, utilizing namespaces, Kata Containers, gVisor, or Firecracker microVMs.
- Manage AI Storage & Fault Tolerance: Orchestrate efficient model weight distribution using OCI Image Volumes and implement Checkpoint/Restore capabilities (via CRIU and NVIDIA cuda-checkpoint) for long-running training fault recovery.
- Enable Distributed Training: Implement robust gang scheduling to prevent deadlocks in tightly-coupled, multi-node training jobs (e.g., MPI, PyTorch) using tools like Volcano, Kueue, or LeaderWorkerSet (LWS).
- Orchestrate Complex Inference: Implement and manage disaggregated AI inference pipelines using frameworks like NVIDIA Grove, coordinating multicomponent deployments (e.g., prefill leaders, decode workers, KV routers) with multilevel autoscaling and explicit startup ordering.
What You'll Bring:- Kubernetes Expertise: Deep technical knowledge of Kubernetes core components, API performance optimization, Dynamic Resource Allocation (DRA), and the custom resource definitions (CRDs) required for advanced scheduling.
- Advanced Scheduling Experience: Proven track record working with AI-specific Kubernetes schedulers and orchestrators such as Kueue, Volcano, Apache YuniKorn, or Run:ai / KAI-Scheduler.
- Hardware & Topology Acumen: Deep understanding of GPU architectures (NVIDIA and AMD) and interconnects, understanding how hardware topology directly impacts training and inference speeds.
- Resource Management Skills: Experience balancing performance and cost using Dominant Resource Fairness (DRF), load-aware scheduling, and bin-packing vs. spread strategies to maximize node vacancy or workload resources.
- Systems Isolation Background: Familiarity with container runtime internals (containerd, runc), rootless containers, and security contexts to manage blast radiuses in shared AI infrastructure.
- AI/ML Framework Knowledge: Strong understanding of modern LLM serving architectures, prefill-decode disaggregation, and engines like vLLM, Triton, or SGLang.
- Observability Proficiency: Experience tracking deep infrastructure and inference metrics, including Time To First Token (TTFT), Time Per Output Token (TPOT), GPU memory pressure, and identifying hardware failures like XID errors.
Compensation Range: - $191,200.00 - $239,000.00
*This is a hybrid role
#LI-Hybrid
