Performance & Systems Engineer Remote / Hybrid · →

The seat sits between the models we serve, the clouds they run on, and the low-level path a request actually takes — inference engines, orchestration, the agent execution stack. Walk the critical path from an agentic loop to the inference pipeline and the container that holds it, and name the hop that is actually expensive. Build the telemetry and profiling the rest of the team can run: latency, memory bandwidth, compute throughput, GPU utilisation, serving cost on the clusters that already exist. Land architecture changes with the people who train, serve, and productise the models — fewer wasted tokens, tighter context handling, a topology that does not multiply RPCs for sport. Tune serve-side workloads: hardware utilisation, context pipelines, microservice shape. The number that moves is the one an operator can feel or a finance lead can read. What you will do Walk the critical path from an agentic loop to the inference pipeline and the container that holds it, and name the hop that is actually expensive. Build the telemetry and profiling the rest of the team can run: latency, memory bandwidth, compute throughput, GPU utilisation, serving cost on the clusters that already exist. Land architecture changes with the people who train, serve, and productise the models — fewer wasted tokens, tighter context handling, a topology that does not multiply RPCs for sport. Tune serve-side workloads: hardware utilisation, context pipelines, microservice shape. The number that moves is the one an operator can feel or a finance lead can read. What you bring ML inference or serving platforms and large cloud estates in the same week: Kubernetes, distributed services, GPU orchestration. A habit of chasing latency and memory through a high-throughput system with traces, profiles and benchmarks you can replay. A production codebase that was messy when you arrived, and a win you can point at after — milliseconds or dollars, written down. The trade: raw speed against compute cost, reliability, and the person waiting on the token stream. What you bring ML inference or serving platforms and large cloud estates in the same week: Kubernetes, distributed services, GPU orchestration. A habit of chasing latency and memory through a high-throughput system with traces, profiles and benchmarks you can replay. A production codebase that was messy when you arrived, and a win you can point at after — milliseconds or dollars, written down. The trade: raw speed against compute cost, reliability, and the person waiting on the token stream. Useful, not required CUDA kernels, Triton, vLLM, TensorRT-LLM, or a C++ / Rust runtime you extended rather than wrapped. Agent loops, context-window management, streaming responses — how they fail when the KV cache or the network is the real budget. Open-source work on ML systems, HPC libraries, or cloud-native tooling. Useful, not required CUDA kernels, Triton, vLLM, TensorRT-LLM, or a C++ / Rust runtime you extended rather than wrapped. Agent loops, context-window management, streaming responses — how they fail when the KV cache or the network is the real budget. Open-source work on ML systems, HPC libraries, or cloud-native tooling. The first quarter Audit latency and cost across inference, the network, and the agent loop. Establish one profiling baseline for latency, throughput, and GPU memory. Land a measured reduction in p99 latency or serving cost on a primary workload.