Research Scientist – Multimodal Foundation Models & Visual Reasoning

Description

We are a well-funded, early-stage AI research and product lab focused on building the next generation of frontier multimodal AI models. Founded by former DeepMind researchers—including Andrew Dai, previously an engineering leader on Gemini—our mission is to push the frontier of visual and multimodal intelligence. Our first model, Glimpse, launched publicly in October 2026.

Our team currently consists of 25 world-class scientists and engineers. We raised a $55M seed round led by Altimeter Capital, Menlo Ventures, Striker VC, and NVIDIA. We are tackling foundational problems in artificial intelligence and translating cutting-edge multimodal research into scalable, real-world systems.

About the role

In this role, you will lead ambitious, high-impact research across next-generation model architectures, distributed training methodologies, multimodal visual reasoning, reinforcement learning, and data-centric learning. Backed by dedicated, large-scale compute clusters and deep engineering support, you will formulate first-principles hypotheses and run large-scale empirical experiments. You will collaborate closely with world-class research and engineering peers to push the boundary of visual intelligence and translate core research breakthroughs directly into state-of-the-art foundation models and real-world capabilities.

Responsibilities

• Frontier Research: Drive novel research at the intersection of foundation models, multimodal understanding, complex visual reasoning, and spatial intelligence.
• Architecture & Training Innovations: Conceive and build novel model architectures, pre-training methodologies, and scalable post-training techniques (e.g., preference tuning, reasoning-focused RL) to systematically advance model capabilities.
• Large-Scale Experimentation: Design, execute, and analyze large-scale distributed training experiments on modern accelerator clusters to push the frontier of model performance.
• Data-Centric Intelligence: Pioneer innovative approaches to multimodal data curation, synthetic data generation, automated filtering, and rigorous benchmark evaluation.
• Breakthrough to Capability: Collaborate with cross-functional research and systems engineering teams to turn fundamental algorithmic breakthroughs into robust, scalable foundation model capabilities deployed at scale.
• What we are looking for
• Ph.D. or equivalent industry research track record in Machine Learning, Computer Vision, Natural Language Processing, or a related quantitative discipline.
• Rigorous foundational understanding of deep learning principles, transformer/multimodal architectures, and modern large-scale training dynamics.
• Deep proficiency in Python and modern deep learning frameworks (PyTorch or JAX).
• Demonstrated ability to independently define research agendas, design rigorous empirical experiments, and solve open-ended, frontier-level problems.

Requirements

• Specialized Domain Expertise: Deep theoretical and practical expertise in at least one of the following: foundation models, multimodal reasoning, visual/spatial intelligence, reinforcement learning, or post-training
• Proven Research Impact: Strong track record demonstrated through first-author publications at top-tier venues (e.g. NeurIPS, ICML, ICLR, CVPR, ICCV, ECCV) or key technical contributions to high-profile foundation model releases.
• Distributed Systems & Scale: Hands-on experience training, profiling, and scaling models across large distributed GPU or TPU clusters (e.g., FSDP, Megatron-LM, DeepSpeed, Ray).
• Data & Evaluation: Demonstrated experience developing large-scale multimodal data pipelines, synthetic reasoning trace generation, or designing rigorous evaluation suites for frontier models.