andromeda - HPC Architect

• Deep HPC experience: you've designed, built, or operated GPU clusters at meaningful scale and understand what makes them fast, stable, and debuggable. • Strong fabric knowledge, ideally both InfiniBand and RoCE. You can evaluate a topology, interpret fabric diagnostics, and identify why a fabric underperforms, not just that it does. • Experience with distributed orchestration and the HPC software stack: Slurm, Kubernetes, OpenMPI or equivalent, and the Linux systems engineering underneath all of it. • Data-center literacy: power, cooling, cabling, and physical-layer realities. You can walk a provider's facility and know what questions to ask. • Benchmarking judgment. You know which numbers matter for large-scale training workloads, how providers game them, and how to design tests that can't be gamed. • The ability to write standards others can build against: precise, testable, and usable by a provider's engineering team without you in the room. • Credibility in the room with external engineering teams, including the ability to deliver a failing grade to a provider who wants your business, and keep the relationship intact. • Comfort with ambiguity. The qualification framework you'll apply mostly doesn't exist yet; you'll write it. • Strong Candidates May Have • Experience inside a neocloud, hyperscaler, colocation, or data-center provider. • NVIDIA data-center GPU depth: DGX/HGX platforms, NVLink/NVSwitch, GPU health and RAS behavior at fleet scale. • Experience supporting AI research labs or other large-scale training customers, and familiarity with what their workloads punish: stragglers, fabric jitter, storage stalls. • Background in cluster acceptance testing or site bring-up. Ideally you've taken a cluster from delivery to production before. • What Success Looks Like • Within your first year, you'll have delivered: • A qualification bar that's written down and trusted. Andromeda's acceptance criteria, benchmark suite, and quality thresholds exist as documents and tooling — not tribal knowledge. SREs, procurement, and providers all work from the same bar, and a passing grade from you means the cluster performs in production. • Providers who build to our standards before we ever test them. Prospective providers get our technical requirements upfront and arrive closer to qualified, because the standards are clear enough to engineer against. Time-from-sourcing-to-qualified drops with each provider. • Onboarding that's technically boring. Clusters that join the network have been validated the same way every time, and the surprises that used to appear in a customer's training run get caught in acceptance instead. • A technical relationship providers value. Provider engineering teams treat you as the authority on what Andromeda needs and the first call when they're planning a build-out so that we hear about architecture decisions early enough to influence them. • A procurement partnership that changes what we buy. Procurement's provider pipeline reflects your technical due diligence, and deals are shaped by a realistic view of remediation cost.