Principal ML Performance Engineer (GPU Optimization)

Sponsor Finder; Boston; On-site; GBP 120,000 - 180,000; Full time

Job summary

Proxima is seeking a Principal ML Performance Engineer to optimize training and inference for state-of-the-art models. You will profile PyTorch, write custom kernels in CUDA/Triton, and leverage compilers like torch.compile, TensorRT, and XLA to maximize throughput on large GPU clusters.

You\'ll scale distributed training across 32–64 nodes on GCP, manage memory scaling for large complexes, and build benchmarks and profiling tools for the research team.

Qualifications

• 6+ years in ML systems, HPC, or performance engineering.
• Deep knowledge of PyTorch internals and profiling bottlenecks.
• Experience with CUDA and Triton; reading Nsight outputs.
• Distributed multi-node training experience.
• Strong proficiency in Python and C++.
• Ability to name a model they made materially faster and quantify the improvement.

Responsibilities

• Profile and optimize training and inference for structural and generative models.
• Write and tune custom kernels (CUDA, Triton) and use compilers (torch.compile, TensorRT, XLA).
• Scale distributed training across 32–64 nodes with advanced parallelism and mixed precision.
• Reduce inference cost by memory scaling and throughput improvements for large GPU workloads.
• Manage GPU cluster efficiency on cloud, including scheduling and cost reporting.
• Develop benchmarks and profiling tools for the research team.

Skills

ML systems

HPC

Performance engineering

PyTorch internals

CUDA

Triton

Python

C++

Distributed training

Education

BS/MS/PhD in CS/EE or related field

Tools

Nsight

TensorRT

XLA

torch.compile

Job description

Principal ML Performance Engineer (GPU Optimization)

About Proxima

Proxima is a frontier AI and data generation company discovering the next generation of proximity therapeutics by making protein interactions programmable. Our platform brings together foundation-model machine learning, a scalable data generation engine, and a partnership track record exceeding $5B in collaborations across the world's leading biopharma and tech organizations. We've recently closed an oversubscribed seed round with an elite group of VCs including DCVC, NVIDIA's NVentures, AIX, Yosemite among others.

Neo-1 is our all-atom foundation model that combines state-of-the-art structure prediction and molecular generation in a single system. Neo-1 enables rapid exploration of chemical and structural space for high value, previously intractable targets, and in particular unlocks small molecule proximity therapeutics like molecular glues with AI for the first time.

In parallel, we are developing an advanced structural interactomics platform built on proprietary XLMS technology and a lab equipped with next-generation mass spectrometry instrumentation. This platform produces proteome-scale maps of protein interactions and helps identify small molecules that modulate proximity. Together with Neo-1, it creates an integrated system capable of co-folding protein complexes while generating candidate small molecules to influence those interactions. Proximity-based therapeutics represent one of the most promising frontiers in modern drug discovery with the potential to treat previously intractable diseases and target 'undruggable' proteins. We're building the tech and the team to make that happen. Come join us.

What you'll do

• Profile and optimize training and inference for structural and generative models, including transformers, diffusion, and geometric deep learning

• Write and tune custom kernels (CUDA, Triton) and use compilers (torch.compile, TensorRT, XLA) when beneficial

• Scale distributed training across 32-64 nodes, employing FSDP, DeepSpeed, tensor and pipeline parallelism, and mixed precision

• Reduce inference cost by optimizing memory scaling for large complexes, improving diffusion sampling efficiency, batching ragged inputs, and maximizing throughput across up to 1000 GPUs

• Manage GPU cluster efficiency on GCP, focusing on scheduling, utilization, spot strategy, and cost reporting

• Develop benchmarks and profiling tools for the research team

What we need

• Minimum of 6+ years experience in ML systems, HPC, or performance engineering, with a BS/MS/PhD in CS, EE, or related field

• Demonstrated ability to set technical direction beyond coding: selecting infrastructure, influencing research teams, and mentoring engineers

• Deep knowledge of PyTorch internals with hands-on experience profiling and fixing real bottlenecks

• Experience with CUDA and Triton, skilled at reading Nsight output, and strong understanding of memory bandwidth and occupancy

• Experience with distributed training at multi-node scale

• Strong proficiency in Python and C++

• Able to name a model they made materially faster and quantify the improvement

Nice to haves

• Experience in geometric deep learning, equivariant networks, or protein structure models such as AlphaFold, ESM, or RFdiffusion

• Experience writing kernels for structure-model primitives, including triangle attention, triangle multiplicative updates, cuEquivariance, or FlashAttention for pair bias

• Experience orchestrating large batch inference and managing Kubernetes GPU scheduling