Staff Research Software Engineer: Open ML Infra & RL Training

Reflection AI Ltd; New York (NY); On-site; USD 180,000 - 260,000; Full time

Benefits offered by this job

Top-tier compensation & equity

Stock options

Health & wellness benefits

Meals provided

Generous parental leave

Unlimited vacation (US) / 30 days (UK)

Visa sponsorship

Team off-sites

Job summary

Reflection AI Ltd. is seeking an engineer to architect and optimize the core training infrastructure powering our frontier AI models.

You will work with researchers to turn new ideas into reliable, scalable training systems capable of handling RL training loops, distributed GPUs, and vast data pipelines. You will collaborate across teams to implement techniques, ensure numerical stability, and reduce bottlenecks, delivering production-ready training stacks that support rapid iteration at massive

Qualifications

• Strong software engineer with ML literacy.
• Ability to translate research into production-grade infrastructure.
• Experience in distributed training or data infrastructure is highly relevant.

Responsibilities

• Design and optimize large-scale training loops and data pipelines.
• Implement state-of-the-art techniques ensuring numerical stability and efficiency.
• Build tooling to launch, monitor, and reproduce complex experiments.
• Diagnose bottlenecks across the training stack (GPU memory, comms, dataloader).
• Translate research prototypes into production-grade infrastructure.

Skills

Strong software engineering

ML concepts & research-to-prod

Distributed systems mindset

Tools

PyTorch

JAX

Megatron-style stacks

Triton

Ray

Kubernetes

Slurm

NCCL

RDMA

Job description

Reflection AI Ltd. is seeking an engineer to architect and optimize the core training infrastructure powering our frontier AI models.

You will work with researchers to turn new ideas into reliable, scalable training systems capable of handling RL training loops, distributed GPUs, and vast data pipelines. You will collaborate across teams to implement techniques, ensure numerical stability, and reduce bottlenecks, delivering production-ready training stacks that support rapid iteration at massive