Senior ML Infrastructure Engineer — Frontier RL & LLM Training

Preference Model; San Francisco (CA); On-site; USD 200,000 - 350,000; Full time

Benefits offered by this job

Health, vision, dental benefits

401K match

Lunch provided onsite

Weekly snack orders

Visa sponsorship & relocation support

Job summary

Preference Model in San Francisco seeks a Senior ML Infrastructure Engineer to design, build, and scale the infrastructure that powers post-training research on in-house RL environments. You will develop core ML framework primitives and tooling to accelerate reproducible experimentation and reduce time from idea to result.

Work closely with Research Engineers to translate research needs into scalable infra while tackling distributed systems challenges, cloud platforms, and high-throughput

Qualifications

• Proven experience building production-grade ML infrastructure.
• Hands-on with distributed training and high-throughput systems.
• Solid experience with cloud platforms and container orchestration.

Responsibilities

• Design, build, and scale compute, scheduling, and data infra for post-training research.
• Develop core ML framework primitives and internal tooling for reproducible experiments.
• Create evaluation, logging, and deployment tooling to catch failures early.
• Collaborate with Research Engineers to translate needs into scalable infra

Skills

LLM inference infrastructure

Distributed systems

Kubernetes

AWS/GCP

PyTorch/JAX

RL training frameworks

Tools

vLLM

Megatron

SGLang

Slime

veRL

Ray Train

SkyRL

Job description

Preference Model in San Francisco seeks a Senior ML Infrastructure Engineer to design, build, and scale the infrastructure that powers post-training research on in-house RL environments. You will develop core ML framework primitives and tooling to accelerate reproducible experimentation and reduce time from idea to result.

Work closely with Research Engineers to translate research needs into scalable infra while tackling distributed systems challenges, cloud platforms, and high-throughput