SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
d-Matrix is building software and hardware for generative AI inference at scale. The Inference Cloud Engineering team develops deployment and orchestration technology for d-Matrix's inference services, including an internal cloud for model testing and tuning, and a customer-facing cloud with self-service inference endpoints.
As Senior Staff Inference Performance Engineer, you will own end-to-end performance of deployed workloads across the full serving path—from request arrival through result delivery. Your mandate is that workloads on d-Matrix's orchestrated platform perform at least as well as on bare metal. Current workloads include speculative decoding and attention-FFN disaggregation across d-Matrix Corsair accelerators and GPUs, with more heterogeneous architectures planned.
You will diagnose performance problems across orchestration (scheduling, placement, networking, autoscaling) and inference internals (batching, KV-cache, prefill/decode, serving framework, driver, runtime, interconnect, host OS, NUMA, accelerator). You will isolate each problem to a specific layer and owner, decide whether to fix it directly, route it to the owning team, or both, and write minimal, reproducible problem statements so owning teams can act independently. You will protect throughput, latency, and stability targets by finding and closing gaps against bare metal, assess scaling requirements from single deployments to fleet operation, and capture workload node requirements in orchestration configuration. You will build lightweight telemetry and benchmarks to isolate or hand off problems, and turn recurring investigations into playbooks while mentoring engineers in the profiling-first method.
The role requires deep systems fundamentals and the ability to work across unfamiliar, sparsely documented stacks. You will use profiling to find performance overhead, work with Kubernetes for performance-sensitive workloads, and understand LLM inference serving internals. Clear technical writing and sound judgment about ownership are essential—you will hand off problems that are not yours to fix with evidence the owning team can act on.
REQUIREMENTS:
- 8+ years of hands-on performance or systems engineering, including resolving hard regressions that crossed multiple layers of a stack
- Track record of finding performance overhead through profiling in unfamiliar, sparsely documented systems (valued more than experience with specific vendor stacks or accelerators)
- Deep systems fundamentals: Linux internals, NUMA, CPU pinning and cpusets, PCIe, memory and cache behavior, network paths
- Experience running performance-sensitive workloads on Kubernetes, identifying when scheduler defaults, network overlays, or orchestration abstractions cost latency or throughput, and closing the gap
- Working knowledge of LLM inference serving: batching, KV-cache, prefill/decode, quantization, speculative decoding
- Proficiency in Python and C/C++, sufficient to instrument, debug, and patch code across the stack
- Clear technical writing that hardware, firmware, or platform teams can act on without context
- Sound judgment about ownership with examples of correctly handing off problems that were not yours to fix
- Fluency with AI and agentic tools for investigation, code search, log triage, and drafting instrumentation, with discipline to verify output against real data
PREFERRED:
- Experience with LLM serving frameworks (vLLM, SGLang, TensorRT-LLM)
- Hands-on experience building or tuning speculative decoding, disaggregated serving, or heterogeneous serving architectures
- Experience tuning KV-cache management, request and batch scheduling, or weight prefetch and residency in LLM serving systems
- GPU programming and profiling experience (CUDA, Nsight, NCCL)
- Experience with pre-production or bring-up hardware, or organizations where root causes routinely cross team boundaries
- Rust experience
- Startup experience or comfort in environments where priorities shift quickly
