Research Engineer, Post-Training Inference

Together AI; San Francisco (CA); On-site; USD 180,000 - 280,000; Full time

Job summary

Together AI in San Francisco is seeking a Research Engineer for Post-Training Inference to build a platform that lets developers tailor open-source models with their data. You will work across post-training, fine-tuning, RL, and evaluation to production serving and optimize the inference engine for large-scale experiments.

You'll collaborate with product, research, and engineering teams to keep the API reliable, performant, and deeply integrated into our infrastructure, advancing open-source AI

Qualifications

• 2+ years of experience building and deploying ML-based services in a production environment.
• Hands-on experience with modern inference engines, such as SGLang, vLLM, and TensorRT-LLM.
• Familiar with the latest methods for fine-tuning LLMs and other AI models.
• Strong software engineering background in Python or Go.
• Stay up to date with the latest advances and trends in ML.

Responsibilities

• Design and build Together’s systems for customizing open-source models.
• Build integrations between the Model Shaping and Inference platforms to ensure a seamless path from post-training to serving production workloads.
• Add features to inference engines for large-scale post-training experiments, including optimizations for RL workloads.
• Make sure the service is stable and robust, participating in an on-call rotation and ensuring 24/7 availability of our platform.

Skills

ML in prod

Python/Go

System design

Open-source ML

Tools

SGLang

vLLM

TensorRT-LLM

CUDA

Job description

Research Engineer, Post-Training Inference

About the role

The Model Shaping team at Together AI works on products and research focused on tailoring open foundation models to downstream applications. We build services that enable machine learning developers to choose the best models for their tasks and further improve these models using domain-specific data. In addition, we develop new methods for more efficient model training and evaluation, drawing inspiration from a broad range of ideas across machine learning, natural language processing, and ML systems.

As a Research Engineer within Model Shaping, you will develop a platform that enables users to customize open-source models with their own data. Working across the training and inference stacks, you will build and improve our Fine-Tuning, Reinforcement Learning, and Evaluation services – from ensuring a seamless path from post-training to production serving, to optimizing the inference engine for RL training workloads. You will collaborate closely with our product, research, and engineering teams to keep the API reliable, performant, and well integrated into the company's technical infrastructure. Above all, you will help build the foundational layer of the open-source AI ecosystem, enabling developers around the world to efficiently create high-quality models tailored to their specific applications.

Responsibilities

• Design and build Together’s systems for customizing open-source models
• Build integrations between the Model Shaping and Inference platforms to ensure a seamless path from post-training to serving production workloads
• Add features to inference engines for large-scale post-training experiments, including optimizations for RL workloads
• Make sure the service is stable and robust, participating in an on-call rotation and ensuring 24/7 availability of our platform

Requirements

• Have 2+ years of experience building and deploying machine learning-based services in a production environment
• Have hands-on experience with modern inference engines, such as SGLang, vLLM, and TensorRT-LLM
• Are familiar with the latest methods for fine-tuning LLMs and other AI models
• Have a strong software engineering background in Python or Go
• Stay up to date with the latest advances and trends in the machine learning community

Experience in any of the following will make you stand out

• Serving low-precision (FP4/FP8) models, multiple LoRA adapters within one model instance (Multi-LoRA), or models distributed across several GPU nodes
• Optimizing the performance of RL training workloads
• Developing CUDA/Triton/C