1 day ago
Berlin, Germany or Cambridge, United KingdomMid Level
Responsibilities
- Profile large-scale AI training and inference workloads to identify bottlenecks in Model FLOP Utilization, memory bandwidth, and interconnect performance.
- Optimize NVIDIA GPU kernels, collective communication layers, and model-to-hardware mappings.
- Implement quantization, model pruning, and distillation to reduce inference latency.
- Streamline data pipelines so high-performance accelerators remain fully utilized.
- Build observability tools and real-time dashboards for training execution, system metrics, cluster uptime, and compute utilization.
- Monitor accelerator usage, prevent idle time and memory leaks, and manage compute efficiency and cost.
- Develop scaling laws and internal tools to predict compute, memory, and networking behavior as models grow.
- Collaborate with AI researchers on hardware-compatible model architectures and with infrastructure teams on future compute-cluster design.
Requirements
- At least 3 years of experience in performance engineering, site reliability engineering, or low-level software development focused on HPC or AI workloads.
- Expert-level knowledge of GPU architecture and experience optimizing CUDA kernels or using Triton.
- Deep understanding of distributed training frameworks such as PyTorch Distributed, DeepSpeed, or Megatron-LM, along with networking protocols.
- Proficiency with profiling and performance-tracking tools such as Nsight Systems, Nsight Compute, or Weights & Biases.
- Ability to work autonomously in an early-stage, fast-paced startup environment.
- Experience with TPUs or alternative accelerators is preferred.
- Academic background or personal interest in materials science, chemistry, or physics is preferred.
- Knowledge of MLOps practices and deploying models at scale in AWS, GCP, or Azure is preferred.
- Contributions to open-source performance optimization or machine learning libraries are preferred.
Benefits
- Competitive salary and equity in CuspAI.
- 28 days of holiday in Germany, the Netherlands, and the UK, or 21 days in Japan, Singapore, and the US, in addition to local public holidays.
- 26 weeks of fully paid parental leave for primary caregivers and 12 weeks for secondary caregivers.
- Professional development budget.
- Hybrid work based in Cambridge, London, Amsterdam, or Berlin, with an expectation of three office days per week.
- Onboarding requires at least a few weeks in the Cambridge or London office.
- Regular travel to European, regional, and global locations may be required.
- Interdisciplinary work with AI research, computational chemistry, and experimental science teams.
Tech Stack
Categories
About CuspAI
CuspAI builds an AI-driven materials discovery platform that helps R&D teams in chemicals, energy, and manufacturing design and evaluate novel molecules and materials. The company offers software and models accessed via cloud APIs and collaborations to prioritize candidates, guide experiments, and shorten lab cycles. Founded in 2024 and headquartered in Cambridge, it is a privately held startup focused on applying generative and physics-informed AI to sustainability and next‑generation materials problems.