Research Software Engineer: Scalable RL Training Systems

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

Benefits offered by this job

Salary and equity

Stock options

Healthcare

Meals provided

Parental leave

Unlimited PTO

Job summary

Reflection is building open AI infrastructure to turn cutting-edge research into scalable training systems. You will design and optimize core training pipelines, RL loops, and large-scale data processing across thousands of GPUs, working closely with researchers to deliver reliable, production-grade infrastructure.

The role emphasizes performance, numerical stability, and reproducibility, with opportunities to contribute across HPC, distributed systems, and advanced frameworks in a high-autonomy

Qualifications

• Strong software engineer who speaks the language of machine learning.
• PhD is not required, but you know how to implement a research paper.
• You have deep experience in at least one: Distributed Training & Inference or Data Infrastructure.
• You enjoy the boundary between ML, distributed systems and HPC.
• You care deeply about performance, numerical stability, and reproducibility.

Responsibilities

• Designing and optimizing large-scale training loops and data pipelines.
• Implementing state-of-the-art techniques with numerical stability and efficiency.
• Building internal tools for launching, monitoring, and reproducing experiments.
• Diagnosing bottlenecks across the training stack (GPU memory, communication, dataloader stalls).
• Translating research prototypes into production-grade infrastructure.

Skills

Distributed Training

Data Infrastructure

HPC

Numerical Stability

GPU Optimizations

Tools

PyTorch

JAX

Megatron

Triton

Ray

Job description

Reflection is building open AI infrastructure to turn cutting-edge research into scalable training systems. You will design and optimize core training pipelines, RL loops, and large-scale data processing across thousands of GPUs, working closely with researchers to deliver reliable, production-grade infrastructure.

The role emphasizes performance, numerical stability, and reproducibility, with opportunities to contribute across HPC, distributed systems, and advanced frameworks in a high-autonomy