Research Engineer / Research Scientist, Vision

We’re looking for research engineers with a strong computer vision background who believe that visual and spatial reasoning are core to fully unlocking the capabilities of LLMs. In this role, you'll work on research, development, and evaluation for state-of-the-art Claude models, with a focus on visual and spatial capabilities. This role is highly collaborative and will touch many aspects of our broader research efforts, taking a full-stack approach across pretraining, RL, and runtime techniques like agentic harnesses. Additionally, you’ll partner with the product org to ensure that the vision improvements you deliver impact Claude’s performance on real-world tasks.

What you'll do:

• Run experiments to evaluate architectural variants, data strategies, and SL and RL techniques to improve Claude’s vision

• Develop and test tools, skills, and agentic infrastructure that enable Claude to reason over visual inputs

• Create evaluations and benchmarks that measure progress on multimodal capabilities across training and deployment

• Work with our product org to find solutions to our most vexing API customer challenges related to vision and spatial reasoning

• Have 7+ years of ML, computer vision, and software engineering experience through industry, academia, or other projects

• Are familiar with the architecture, training, and operation of large vision language models

• Have experience creating and evaluating large synthetic and real-world visual training datasets

• Have experience engaging in systematic prompting, finetuning, or evaluation

• Are results-oriented, with a bias towards flexibility and impact

• Enjoy pair programming and cross-team collaboration

• Care about the societal impacts of your work

• Large-scale pretraining, SL, and RL on language models

• Deep learning research on images, video, or other modalities

• Developing complex agentic systems using LLMs

• High-performance ML systems (GPUs, TPUs, JAX, PyTorch)

• Large-scale ETL and data pipeline development

• Running experiments to determine ideal training datamixes and parameters for a synthetically generated vision dataset

• Finetuning Claude to maximize its performance using a particular set of agent tools/skills

• Building a pipeline to ingest and process a novel source of visual training data

• Designing and running experiments to evaluate the scalability of two architectural variants