Senior Software Engineer (Inference)
Posted Updated
- Senior engineers are area owners who lead designs, raise engineering standards, and deliver measurable improvements to latency, throughput, and reliability across multiple services
- You’ll partner with product, orchestration, and hardware teams to evolve our Kubernetes-native inference platform and meet strict P99 SLAs at scale
- Lead design reviews and drive architecture within the team; decompose multi-service work into clear milestones
- Define and own SLIs/SLOs; ensure post-incident actions land and reliability improves release-over-release
- Implement advanced optimizations (e.g., micro-batch schedulers, speculative decoding, KV-cache reuse) and quantify impact
- Strengthen incident posture: capacity planning, autoscaling policy, graceful degradation, rollback/traffic-shift strategies
- Mentor IC1/IC2 engineers; review cross-team designs and elevate coding/testing standards
- Own an area spanning multiple services and teams (e.g., request routing & adaptive scheduling, cost-per-token analytics, GPU resource isolation)
- Hands-on experience with Kubernetes at production scale, CI/CD, and observability stacks (Prometheus, Grafana, OpenTelemetry)
- ~ 5-8 years industry experience building distributed systems or cloud services
- Optimize end-to-end ML system performance by developing and tuning CUDA kernels, reducing model latency, maximizing compute and memory bandwidth utilization, and leveraging custom accelerators for high-efficiency workloads
- Strong coding in Python or Go (C++ a plus) and deep familiarity with networked systems and performance
- Practical knowledge of inference internals: batching, caching, mixed precision (BF16/FP8), streaming token delivery
- Proven track record improving tail latency (P95/P99) and service reliability through metrics-driven work
- Leading multi-team initiatives or partnering with customers on mission-critical launches
- Experience with CUDA kernels, NCCL/SHARP, RDMA/NUMA, or GPU interconnect topologies
- Contributions to inference frameworks (vLLM, Triton, TensorRT-LLM, Ray Serve, TorchServe)