Employer description & source details
Every model we train runs on infrastructure this role owns. We operate NVIDIA GPU clusters across bare metal and rented capacity, and we're looking for an engineer to join our small research infrastructure team and make that compute fast, reliable, and boring - in the best sense. When the clusters just work, research moves faster. Your impact is measured directly in training throughput and researcher velocity. This is a builder-operator role with real breadth: one week you're writing automation that eliminates a whole class of manual work, the next you're benchmarking a new provider's InfiniBand fabric or on-site bringing new hardware online. You'll have unusual scope and autonomy - we're a lean team where decisions are made by the people closest to the problem. Operate and improve our GPU fleet end to end: provisioning, scheduling, monitoring, upgrades, capacity planning Build automation that keeps the fleet healthy without human intervention — node health checks, automated draining and remediation, burn-in pipelines for new capacity Own the stack beneath the training code: OS images, NVIDIA drivers, CUDA, container runtimes, NCCL, high-speed networking (InfiniBand/RoCE) Run and tune job scheduling (Slurm or similar) so researchers get compute fairly and fast Build and maintain high-performance storage for datasets and checkpoints Hunt down performance problems: stragglers, degraded links, thermal issues, flaky GPUs — and fix the class of problem, not just the instance Evaluate rented GPU capacity: benchmark it, validate it, hold providers to their SLAs Hands-on hardware work when it's needed: racking, cabling, diagnostics, coordinating with datacenter staff and vendors Keep clusters secure by default: access control, network isolation, secrets Have run large-scale Lin
Extracted skills: python, terraform