> Markdown version of [/jobs/ext/4722809-senior-machine-learning-engineer-gpu-optimization](https://www.wearedevelopers.com/jobs/ext/4722809-senior-machine-learning-engineer-gpu-optimization). Every page supports `.md` or `Accept: text/markdown`. Links point to the HTML versions so they work for humans too. Agent guide: [/agents.md](https://www.wearedevelopers.com/agents.md). --- # Senior Machine Learning Engineer, GPU Optimization - **Company:** Boltz - **Location:** London, United Kingdom - **Experience:** Expert - **Contract:** Permanent contract - **Skills:** 3d Models, Computer Clusters, Profiling, Nvidia CUDA, Extract Transform Load (ETL), Python (Programming Language), Machine Learning, Open Source Technology, Pytorch, DeepSpeed, Deep Learning, AI Coding Agents, Kubernetes, Megatron - **Published:** November 15, 2026 - **Apply:** https://www.indeed.com/viewjob?jk=522e57254a197cb0 ## About the Role * Experience in optimizing deep learning training on multi-node GPU clusters (e.g. FSDP, DeepSpeed, Megatron). * Deep GPU performance expertise: profiling (e.g. PyTorch Profiler, Nsight), memory and communication tuning, and kernel work in Triton or CUDA. * Strong PyTorch and Python, with enough ML intuition to judge how an engineering change affects training dynamics and results. * A track record of measurable speedups or reliability gains on real, large training workloads. * You enjoy investigating problems with researchers, rather than taking tickets., * Exposure to structure prediction, diffusion models, or biology and chemistry data. * Experience with Kubernetes and cluster operations; coding agents as a core part of your workflow. * Experience using Boltz models, or contributions to open-source training frameworks or kernel libraries. We know great people don't always tick every box. If this role excites you and you meet most of the minimum requirements, we'd like to hear from you. ## Description As a ML Engineer in GPU Optimization, you will own how fast and how reliably our structure models train. Every hour cut from a training run is another experiment our researchers can run; a near-term goal is a 1.5x speedup on our structure model, with long runs finishing without interruption. You will spend most of your time embedded with a research team (such as pre-training, post-training or biologics), working side by side with research scientists, and the rest with Research Engineering on tools shared across teams. You'll also work with Platform Engineering, who own GPU orchestration across Boltz. This role is ideal for an engineer who loves profiling, can look at a GPU utilization or loss curve and know where to start digging, and wants their speedups to show up in better science. What you'll do * Speed up large-scale training of our structure models, finding and removing bottlenecks in kernels, memory, communication, data loading and parallelism. * Write and tune custom GPU kernels where they pay off (e.g. Triton or CUDA), and get the most from each new GPU generation. * Make long training runs complete without interruption: diagnose failures and degraded runs from throughput, utilization and loss signals, and fix the root cause. * Build reusable profiling and optimization tooling, including agent-assisted, profile-guided optimization, so speedups carry across teams. * Feed what you learn about performance back into how researchers design models and experiments.