Research Scientist

Roboflow•San Francisco, CA

15d•$250,000 - $450,000•Onsite

About The Position

Roboflow is building the visual cortex for AGI. Roboflow Labs is our applied research unit, focused on training open models, building benchmarks grounded in real enterprise use cases, and partnering with frontier labs to deliver necessary training data. The Research Scientist will work at the intersection of our benchmarks, models, and data customers, conducting failure analyses, defining dataset and evaluation specifications, and validating data quality. This role involves publishing research, extending existing benchmarks, and contributing to the research agenda with work grounded in real-world deployments. The scientist will design datasets, evaluations, and RL environments for vision-language, multi-modal, robotics, and coding models; perform failure analyses on partner models; lead research for new benchmarks; empirically validate data quality; contribute to open models like RF-DETR; and collaborate with research scientists at frontier labs.

Requirements

PhD or equivalent research experience in computer vision, multi-modal learning, machine learning, or a closely related field, with a track record of published work or shipped research
Hands-on experience with post-training: fine-tuning, RL, evaluation design, or data curation for large models
Strong engineering fundamentals — you write the code for your own experiments and can hand it to an engineer to productionize
Clear written communication; you can explain a failure mode to a lab researcher and a data spec to an annotator

Nice To Haves

experience at a frontier lab or as a data/evaluation partner to one
experience with vision-language models, robotics learning, or agentic coding evaluations

Responsibilities

Design datasets, evaluations, and RL environments that target measurable gaps in frontier vision-language, multi-modal, robotics, and coding models
Run structured failure analyses on partner labs' models against Roboflow's real-world task distribution and write the remediation specs that say what data will close each gap
Own the research behind new benchmarks that we publish.
Validate data quality empirically — fine-tune, ablate, and measure so that every dataset we sell comes with evidence it works
Contribute to Roboflow's open models (RF-DETR and successors) and represent Labs within the research community
Work directly with research scientists at frontier labs as a technical peer, translating their model roadmaps into concrete data programs