Research Scientist – Multimodal Foundation Models

The role

You will research, develop, and improve vision-language and multimodal foundation-model architectures. The role involves working on model training, post-training, reinforcement learning, and multimodal reasoning.

What you will do

• Research, develop, and improve vision-language models and multimodal foundation-model architectures.
• Design and implement advanced training strategies including supervised fine-tuning, preference optimization, and reinforcement learning.
• Investigate approaches to improve reasoning, generalization, reliability, and training efficiency.
• Develop strategies for data curation, filtering, and synthetic data generation.
• Design experiments, evaluation frameworks, and benchmarks to measure model performance.
• Work with distributed training infrastructure and large-scale datasets.
• Explore emerging approaches in multimodal learning, model alignment, and foundation-model training.

Required qualifications

• Hands-on experience researching, training, fine-tuning, or post-training large-scale foundation models, particularly vision-language or multimodal models.
• Strong understanding of deep learning architectures, optimization techniques, and modern model training methodologies.
• Practical experience with supervised fine-tuning (SFT), preference optimization, or reinforcement learning approaches such as DPO, GRPO, PPO, or RLHF.
• Strong proficiency in Python and PyTorch, JAX, or comparable deep learning frameworks.
• Experience designing, executing, and analyzing large-scale machine learning experiments.
• Familiarity with distributed training systems and GPU-based model development.
• Strong research instincts, analytical problem-solving skills, and the ability to independently drive complex technical projects.
• Experience working with open-weight multimodal foundation models.
• Research publications on top-tier journals/conferences, open-source contributions, or experience within leading AI research organizations.

Nice to have

• Experience with large-scale model post-training and reinforcement learning for reasoning or alignment.
• Familiarity with multimodal data pipelines involving images, video, or other sensory inputs.
• Experience with reward modeling, synthetic data generation, or advanced model evaluation.
• Experience improving model capabilities through innovations in training strategies, data quality, or model architecture.

Before you apply

• We may use AI-assisted tools to help evaluate applications against the requirements of this role.

About the company

Mecka AI is building the data infrastructure layer for robotics and embodied AI. They design and operate global systems for data capture, data labeling, and hardware-enabled workflows used by leading AI labs and robotics companies.