Inference Optimization Lead

Lead, Autonomous GPU Performance Optimization | Remote (US preferred) Most GPU spend is wasted, and usually the software is the reason, not the hardware. My client is a venture-backed, deeply technical startup building a performance optimization platform for AI workloads. Its agent runs in production environments, profiles training and inference workloads, and pinpoints exactly where GPU time is going. One customer recently cut around 30% of GPU time from a training workload, which saved them millions. The platform runs on NVIDIA and AMD today, with TPUs and other accelerators next on the roadmap. It’s self-hosted and works across hardware vendors, which is why customers who won’t let their data leave their network choose it. The role You’ll own the company’s research program in autonomous optimization. Today the product finds the hotspots. Your job is to build the layer that fixes them: a harness of skills, subagents and tooling that takes in profiling data, understands it, and optimizes the code with very little human hand-holding, on any hardware and any inference engine. This is a long-term strategic program, not a single project. You’ll work directly with the two technical founders, set the direction, and can build a team under you as it grows. What you’ll do Turn hands-on GPU and inference optimization expertise into autonomous, agent-driven workflows Break ambiguous problems into measurable steps, write the design docs and drive the work end to end Extend support to new accelerators and inference engines Work with customers’ real production workloads, not toy benchmarks What you’ll bring Hands-on experience optimizing inference or GPU workloads in production, and having done the work yourself is essential Deep familiarity with some of: CUDA, Triton, CUTLASS, ROCm/HIP, vLLM, SGLang, TensorRT-LLM, FlashAttention, KV-cache and batching strategies, ML compilers (XLA, MLIR, TVM, torch.compile) Comfort with profilers such as Nsight Compute/Systems, rocprof and the PyTorch profiler, and with reading performance data to find the real bottleneck A track record of working independently and writing clear technical design docs Bonus: open-source contributions to inference engines or kernel libraries, or publications This role is about getting more performance out of the GPU itself. It isn’t a DevOps, Kubernetes or ML platform role. What’s on offer A founding-level research leadership role with real ownership Competitive base salary Stock options you can exercise early, at a very low current strike price Fully remote, with a preference for the US. You’ll need some flexibility on hours to overlap with founders based in the UK. A quick process: founder screen, a practical design-doc and coding exercise, a technical deep-dive, then an offer, usually within 1-2 weeks