(Staff / Senior Staff) AI Infra Engineer, Inference Infrastructure

About The Role We're looking for a software engineer, reliability-minded but still a builder, to own what happens in production to the Kubernetes-native control plane that provisions and runs our GPU inference fleet. The platform team builds the manifest-driven API and controllers that let the inference team declare what they need, a cluster, a model deployment, a capacity change, and get it reconciled underneath. Your job is to keep that promise true in the real world: triaging InfiniBand issues when prefill and decode nodes stall exchanging KV cache over RDMA, diagnosing hero-model and LoRA-serving faults that don't show up in generic health checks, and extending the automated self-healing system so each new failure mode becomes a permanent fix instead of a recurring page. The core value we're after is keeping the decoupling real even when things break, and keeping the fleet efficient without that efficiency becoming a new source of instability. You'll own SLOs and error budgets for inference serving, lead incident response, and turn what postmortems teach you into actual changes in the controllers and automation, not just updated runbooks. You'll also work the reliability side of capacity efficiency, since a mistuned defragmentation or bin-packing pass is itself a production risk. Strong candidates have hands-on production-reliability experience on large-scale distributed or GPU-serving systems, are comfortable troubleshooting InfiniBand/RDMA and Kubernetes controller/CRD-based platforms, and think about GPU infrastructure, and the tools that keep it reliable, as software to be engineered. A product mindset applies: the on-call experience, the clarity of an alert, and the judgment built into an automated remediation are things you ship and iterate on. Responsibilities Operate the provisioning state machine in production: own the health of the host lifecycle pipeline, discovery, inference bring-up, GPU driver/CUDA validation, health checks, decommission/RMA, triaging stuck or failed state transitions at fleet scale and feeding real failure patterns back into how the states and transitions are designed. Triage and extend automated self-healing: when the node-scanning and remediation platform misses something, a new failure mode, an InfiniBand issue in disaggregated prefill/decode serving, a GPU Xid variant, a hero-model or LoRA-specific fault, diagnose it live, drive the fix to production, and turn it into a new detection/remediation plugin so it's automatic the next time. Own end-to-end reliability of the pipeline: hold the SLOs and error budgets for provisioning and fleet health, lead incident response when idempotency, retry, rollback, or drift assumptions break down in practice, and drive the systemic fix, not just the hotfix. Partner with the inference/ML platform team from the production side: be the one who knows how real cluster topologies, interconnects, and scheduling constraints actually fail in the field, and feed those failure modes back as first-class requirements into the platform's abstractions. Work alongside machine learning researchers, engineers, and product managers to bring our latest technologies into production, enabling advanced research through awesome engineering. Introduce new techniques, tools, and architecture that improve the performance, latency, throughput, and efficiency of our model inference stack. Build tools to give us visibility into our bottlenecks and sources of instability, then design and implement solutions to address the highest-priority issues. Lead reliability cross-functionally: build coalition across infra, ML platform, and product teams around shared reliability priorities, manage stakeholders through incidents and roadmap trade-offs, and represent the reliability point of view in planning so it shapes decisions before they ship, not after an outage. Requirements Experience operating a provisioning or host-lifecycle pipeline in production,