About the role
ABOUT BASETEN Baseten powers mission-critical inference for the world's most dynamic AI companies, like Cursor, Notion, OpenEvidence, Abridge, Clay, Gamma and Writer. By uniting applied AI research, flexible infrastructure, and seamless developer tooling, we enable companies operating at the frontier of AI to bring cutting-edge models into production. We're growing quickly and recently raised our $1.5B Series F , led by Altimeter Capital, Conviction Partners, and Spark Capital. Join us and help build the platform engineers turn to to ship AI products.
THE ROLE Baseten is building its own GPU infrastructure for large-scale inference. As we move into large scale, high-density NVIDIA systems, the hardest failures are intermittent, cross-layer, and difficult to prove: RoCE congestion, InfiniBand stalls, ECN/DCQCN mis-tuning, bad optics, RNIC issues, host kernel stalls, GPU driver problems, and workload symptoms that look like network problems, but are not. We are hiring a Lead Software Engineer to build a first-class observability and root-cause analysis system for GPU fabrics. This is a hard distributed systems problem, not a dashboarding problem. The system will collect high-volume signals from switches, hosts, active probes, and inference services; reduce and correlate them in real time; understand topology and service ownership; and produce actionable diagnosis while an incident is still unfolding. This role sits at the boundary between networking and inference software. RDMA data paths, GPUDirect transfers, prefill/decode disaggregation, KV cache movement, request routing, and workload backpressure can all create fabric symptoms or hide real fabric failures. The goal is to tell an operator, quickly and with evidence, whether an incident is caused by the fabric, host, NIC, GPU, RDMA path, scheduler, or serving layer — and what to do next. EXAMPLE INITIATIVES Real-time telemetry engine — Build the ingestion, reduction, storage, and query path for high-cardinality fabric, host, GPU, and workload telemetry. Service-aware fabric diagnosis — Build collectors, probes, and topology-aware correlation to detect latency, drops, stalls, congestion, bad paths, and degradation. Software-aware RDMA diagnosis — Tie network behavior to RDMA operations, GPUDirect paths, KV cache transfers, prefill/decode disaggregation, and request latency.