Platform Engineer: Self-Serve GPU Cluster Infra

Wwshemi; United States; Remote; USD 180.000 - 260.000; Vollzeit

Zusammenfassung

Perplexity is seeking an experienced platform engineer to own a self-serve GPU compute platform for training and inference workloads. You will design, build, and operate a unified platform that hides infrastructure complexity from inference researchers and engineers.

You will manage GPU clusters across providers, implement scheduling and placement, and ensure reliable, low-latency service delivery for both long-running training jobs and real-time inference workloads.

Qualifikationen

• Deep Kubernetes experience with custom operators, CRDs, and multi-cluster federation.
• Experience managing GPU clusters at scale with NVIDIA hardware and CUDA.
• Experience orchestrating compute across multiple clouds (AWS, GCP, or similar).
• Strong distributed systems fundamentals: scheduling, resource allocation, and fault tolerance.
• Proficiency in infrastructure code using Go, Rust, or C++.

Aufgaben

• Build a self-serve compute platform for training and inference workloads.
• Operate and integrate the GPU fleet across providers.
• Develop scheduling and placement to maximize capacity.
• Support long-running training jobs and production inference services.
• Own Kubernetes orchestration for GPU across multiple clusters.
• Develop fault tolerance, autoscaling, and observability.

Kenntnisse

Kubernetes

Custom CRDs

Multi-cluster federation

Distributed systems

Go / Rust / C++

GPU clusters

NVIDIA CUDA

Multi-cloud orchestration

Networking (InfiniBand/RoCE)

Tools

Go

Rust

C++

Jobbeschreibung

Perplexity is seeking an experienced platform engineer to own a self-serve GPU compute platform for training and inference workloads. You will design, build, and operate a unified platform that hides infrastructure complexity from inference researchers and engineers.

You will manage GPU clusters across providers, implement scheduling and placement, and ensure reliable, low-latency service delivery for both long-running training jobs and real-time inference workloads.