5-7 years of experience in distributed systems or high-performance computing
Deep understanding of LLM inference and related technologies
Hands-on experience with GPU communication technologies
Proficient in Python and C++ or similar languages
Ability to profile and debug across multiple system layers
Experience with Kubernetes for production workloads
Master's or PhD in a relevant field
Responsibilities
Design and implement disaggregated LLM serving architectures
Optimize multi-node inference on modern GPU systems
Build and tune serving systems using advanced frameworks
Profile end-to-end performance metrics and improve latency
Develop benchmark and capacity-planning workflows
Productionize serving on Kubernetes with reliability standards
Diagnose complex distributed system failures
Collaborate with researchers and engineers to launch new models
Benefits
Collaborative work environment in Menlo Park office
Opportunity to work on cutting-edge AI technologies
Engagement with cross-functional teams
Focus on professional development and growth
Access to advanced tools and frameworks for innovation
Full Job Description
ABOUT THE ROLE
We're seeking an inference systems engineer to build the distributed serving and high-performance networking layer behind our large language models. You will own the path from model server to GPU fabric: prefill/decode disaggregation, KV-cache transfer, multi-node execution, and the observability and benchmarks needed to make those systems reliable in production. This role sits in the Inference team within the company's Model team and partners closely with Model, Infrastructure, and Application engineering. WHAT YOU'LL DO
Design, implement, and operate disaggregated LLM serving architectures, including independent prefill and decode pools, KV-cache transfer, routing, batching, and failure recovery.
Optimize multi-node inference across modern GPU systems using NVLink, NVSwitch, NVLS, InfiniBand, RoCEv2, GPUDirect RDMA, NCCL, UCX, and related communication paths.
Build and tune serving systems using SGLang, vLLM, NVIDIA Dynamo, TensorRT-LLM, or comparable frameworks.
Profile end-to-end performance across compute, memory, network, and storage; improve time to first token, inter-token latency, throughput, tail latency, and cost per token.
Develop repeatable benchmark and capacity-planning workflows across model architectures, GPU types, parallelism strategies, and concurrency levels.
Productionize serving on Kubernetes with health checks, autoscaling, safe rollouts, metrics, tracing, and actionable diagnostics.
Diagnose complex distributed failures such as collective timeouts, topology mismatches, packet loss, congestion, KV-transfer stalls, GPU OOMs, and uneven load.
Partner with model researchers and platform engineers to launch new models, serving features, and hardware generations safely.
LOCATION REQUIREMENT
We believe the best ideas happen together. To support fast collaboration and a strong team culture, this role is expected to be in our Menlo Park office five days a week, unless otherwise specified. WHAT YOU BRING Must-Have:
Production experience building or operating distributed systems, high-performance computing systems, or large-scale ML inference platforms.
Strong understanding of LLM inference, including tensor/pipeline/data parallelism, continuous batching, KV-cache management, and prefill/decode behavior.
Hands-on experience with GPU communication and networking technologies such as NCCL, NVLink/NVSwitch, InfiniBand, RoCE, RDMA, UCX, or equivalent systems.
Strong Python skills and working proficiency in C++ or another systems language.
Ability to profile and debug performance across application, runtime, kernel, network, and infrastructure layers.
Experience deploying production workloads on Kubernetes and operating them with clear reliability and observability standards.
Clear written and verbal communication across research, infrastructure, and product-facing teams.
Master's or PhD Required
Nice-to-Have:
Direct experience with disaggregated serving, KV-cache transfer, NVIDIA Dynamo/NIXL, SGLang, vLLM, or TensorRT-LLM.
Experience with NVLS, SHARP, GPUDirect RDMA, UCX, RDMA congestion control, or GPU-cluster topology optimization.
CUDA, Triton, custom kernel, or low-level GPU performance experience.
Experience with speculative decoding, multi-LoRA serving, quantization, or cache-aware routing.
Contributions to open-source inference, networking, or distributed-systems projects.
Experience benchmarking new GPU platforms and turning results into production architecture decisions.
REFERENCES
Polaris: A Safety-focused LLM Constellation Architecture for Healthcare - https://arxiv.org/abs/[redacted]3
