Summary
What you’ll impact
The role is responsible for end‑to‑end health, reliability and automation of the organization’s massive GPU compute fleet. You will build metrics, alerting, repair pipelines, GPU qualification platforms and low‑level tooling such as Redfish/BMC to ensure scalable operation of thousands of GPUs across Kubernetes and bare‑metal environments.
Responsibilities
What you'll do
- Build the repair pipeline that keeps pace with a fleet of 10s to 100s of GWs: at our scale, a GPU failure isn't a ticket. It's a throughput problem. We're building the automation that takes a chip from fault detection through triage, RMA, and return to service without human intervention.
- Qualify every new GPU generation inside a 6-month build window: our platform covers burn-in, performance baselining, and NPI execution. It has to define "production-ready" before a site goes live, not after. New hardware gets certified at speeds unheard of in the industry.
- Migrate live compute at construction speed: we're converting clusters across production sites simultaneously, bringing new sites online, and making Kubernetes-orchestrated bare metal sustainable at the pace we're building – multiple GW annually.
- See and own the entire fleet in real time, at any scale: build the observability and orchestration layer that makes hyperscale AI compute actually operable. Debug, tune, and performance-test infrastructure that grows by another site every few months.
- Own compute fleet health end to end. Build the metrics pipelines, alerting, and unified health view that tell you the true state of every GPU in production — across Kubernetes-orchestrated workloads and bare metal, at scale.
- Turn deployment/repair into a pipeline, not a procedure. Build and own the automation that takes a compute failure from detection through triage, parts management, and return to service. No one-off scripts, no heroics.
- Design and expand the GPU qualification platform. Burn-in, performance baselining, and NPI execution for every new GPU generation. You define what "good" looks like before hardware goes into production.
- Own Redfish and BMC tooling. Firmware-level telemetry, log collection at fleet scale, and the low-level access layer that repair automation and health tooling depend on.
- Own end-to-end reliability, scalability, and operation of the compute fleet at-scale. Fluidstack is building one of the largest GPU fleets in the world and that can only be accomplished with aggressive automation, tooling, and incident discipline.
Requirements
What you’ll bring
- You treat toil as a bug. Manual steps in a repair workflow are a backlog item, not a job description.
- You have an instinct for hardware. You're comfortable reasoning about failure modes at the firmware and silicon level, not just the software stack above it.
- You move toward ambiguity, not away from it. You walk into the fog, build the map, and explain it to everyone else.
- You learn at a steep slope. You reach real competence in an unfamiliar domain fast. We value this over existing expertise.
- You carry a pager without flinching. You run the incident, write the postmortem, fix the systemic cause, and move on.
- You're fluent with AI tooling. LLM APIs, MCP servers, and agentic frameworks, and you drive Claude Code, Cursor, or similar every day.
- You've shipped production automation that other teams depend on, and you're comfortable in any language using AI coding tools.
- Bonus: Hardware lifecycle management and RMA automation. BMC/Redfish or IPMI tooling. GPU qualification or burn-in frameworks. Workflow and orchestration engines (Temporal, Cadence). Metrics and alerting pipelines (Prometheus, Grafana). Go or Python...