Senior Machine Learning Researcher, Post-Training & Inference Optimization

Thurn Partners Ltd; New York (NY); On-site; USD 180,000 - 300,000; Full time

Job summary

Thurn Partners Ltd in New York seeks a post-training ML researcher to adapt large pretrained models for financial markets, handling noisy data and distributions that shift daily while meeting tight latency budgets.

You will work with production teams to deploy models, evaluate against real market outcomes, and optimize GPU performance with C++/Python, CUDA and bespoke kernels for low-latency inference.

Qualifications

• Degree in quantitative field with GPA ≥ 3.5/4.0
• Industry ML experience with hands-on post-training work
• Production-quality Python and C++
• GPU knowledge of memory hierarchy and how to diagnose a slow kernel
• Eligibility to work in the US

Responsibilities

• Fine-tune, distil and align large models for domain-specific tasks
• Build evaluation that measures what matters, not what is easy to measure
• Quantize, prune and optimize inference where latency carries a direct cost
• Write and profile custom GPU kernels when standard ops become the bottleneck
• Take research into production with engineering and trading teams

Skills

ML experience

Leadership experience

Education

Bachelor's degree in quantitative field (GPA ≥ 3.5/4.0)

Master's or PhD in ML/CS/Math/Stats/Physics

Tools

Python

C++

CUDA

Triton

Nsight

NCCL

Job description

A top-tier global trading firm is building out its post-training research. The mandate is to take large pretrained models and adapt them to financial markets. Those models have to deal with noisy data, distributions that shift daily, and competitors who adapt to any edge. They also have to run within tight latency budgets. Performance is measured against real market outcomes rather than benchmarks or human raters. The firm trades at a scale few can match, so small model improvements have a measurable effect within days. You will join a small group of researchers whose work goes straight into production.

THE ROLE:

• Fine-tune, distil and align large models for domain-specific tasks
• Build evaluation that measures what matters, not what is easy to measure
• Quantize, prune and optimize inference where latency carries a direct cost
• Write and profile custom GPU kernels when standard ops become the bottleneck
• Take research into production with engineering and trading teams

STACK:

• GPU: CUDA, Triton, Nsight, NCCL
• C++17/20, pybind11
• FSDP, DeepSpeed or Megatron-style parallelism

REQUIRED:

• A degree in a quantitative field from a leading university, with a minimum GPA of 3.5/4.0
• Industry experience in ML roles, with hands-on post-training work
• Production-quality Python and C++
• GPU knowledge below the API: memory hierarchy, occupancy, and how to diagnose a slow kernel
• Eligibility to work in the US

PREFERRED:

• A PhD or Master's in machine learning, computer science, mathematics, statistics or physics
• Leadership experience within ML roles, such as technical lead, mentoring or owning a workstream

HOW YOU WORK:

• You start on problems without waiting to be asked, and you know when to bring in your lead
• You treat feedback as data and change course when it is right
• You want to be the least experienced person in the room some days, because that is where the learning is