Low-Latency ML Performance Engineer for GPU Systems

Fintal Partners; New York (NY); On-site; USD 180,000 - 240,000; Full time

Job summary

Fintal Partners is seeking a Machine Learning Performance Engineer to join a small, highly technical team. You will write and optimize CUDA kernels, build scalable training and inference pipelines, and push performance across GPU and software stacks.

You’ll work with researchers and traders to turn model ideas into fast production systems, exploring deep learning frameworks and GPU architectures to extract maximum throughput and min latency in live trading workloads.

Qualifications

• Hands-on CUDA kernel development with Tensor Cores, cooperative groups, CUDA Graphs, and warp-level intrinsics.
• Strong C++ and Python.
• Deep understanding of computer architecture, including memory hierarchy and GPU execution.
• Expertise in the internals of deep learning frameworks such as PyTorch, JAX, or TensorFlow.
• Experience building or optimizing high-performance deep learning systems in any domain.

Responsibilities

• Write and optimize custom CUDA kernels for latency- and throughput-critical deep learning workloads.
• Build scalable, robust training and inference pipelines.
• Identify and eliminate performance bottlenecks across the GPU, memory hierarchy, and software stack.
• Dig into the internals of open-source deep learning frameworks (PyTorch, JAX) and extend them where needed.
• Work directly with researchers and traders to turn model ideas into fast production systems.
• Develop a deep understanding of the trading systems your work supports.

Skills

CUDA kernel development

C++

Python

Tensor Cores

CUDA Graphs

cooperative groups

warp-level intrinsics

deep learning frameworks internals

high-performance DL systems

Tools

PyTorch

JAX

TensorFlow

Triton

cuDNN

cuBLAS

XLA

Job description

Fintal Partners is seeking a Machine Learning Performance Engineer to join a small, highly technical team. You will write and optimize CUDA kernels, build scalable training and inference pipelines, and push performance across GPU and software stacks.

You’ll work with researchers and traders to turn model ideas into fast production systems, exploring deep learning frameworks and GPU architectures to extract maximum throughput and min latency in live trading workloads.