Hands-on experience with large multimodal models, including training and fine-tuning.
Strong deep learning fundamentals with proficiency in PyTorch or JAX.
Experience designing and executing large-scale experiments.
Familiarity with GPU-based development and distributed training infrastructure.
Research instincts paired with engineering discipline for robust system building.
Responsibilities
Build and enhance vision-language and video models for egocentric and manipulation footage.
Develop model pipelines to identify capture defects and task errors, validating against human audits.
Improve model reliability through supervised fine-tuning and reinforcement learning techniques.
Create pipelines for video annotation, including captioning and temporal grounding at scale.
Conduct large experiments on distributed GPU infrastructure and analyze results for production implementation.
Benefits
Opportunity to work on cutting-edge vision-language and video models.
Ownership of projects from data selection to production.
Collaborative environment that values research and engineering skills.
Encouragement for diverse applicants, valuing capability over a perfect checklist match.
Full Job Description
The role
Every capture tells a story: what someone did, when they did it and whether they completed the task. At Mecka, you'll build vision-language and video models that turn that footage into useful annotations and catch quality issues across more video than a human team could review.
You'll own the work from data selection and training through evaluation and production. We're looking for someone who brings strong research instincts and the engineering skills to make models reliable at scale-and who wants to understand where they fail as much as where they succeed. What you'll do
Build models that understand the footage. You'll develop and improve vision-language and video models for egocentric and manipulation footage.
Catch problems across what we collect. You'll build model pipelines that identify capture defects, missing coverage and task errors, then check their performance against human audits.
Make models more reliable. You'll use supervised fine-tuning, preference optimization and reinforcement learning to improve reliability and reasoning.
Turn video into useful annotations. You'll build pipelines for captioning, temporal grounding and video annotation at scale.
Take experiments through to production. You'll run large experiments on distributed GPU infrastructure, analyze the results and turn successful approaches into production systems.
What you'll bring
Hands-on experience with large multimodal models. You've trained, fine-tuned or post-trained vision-language or multimodal models.
Strong deep learning fundamentals. You're comfortable building and training models with PyTorch or JAX.
Experience running large experiments. You can design an experiment, see it through and draw sound conclusions from the results.
Familiarity with distributed training. You know your way around GPU-based development and distributed training infrastructure.
Research instincts and engineering discipline. You ask good questions, test your assumptions and build clean systems that hold up beyond the experiment.
Nice to have
Experience in one or more of these areas would be valuable:
Footage quality control: Using vision-language and video models to identify defects and gaps in captured data.
Video understanding and annotation: Structured captioning, temporal grounding and scene understanding.
Training and post-training: Reward modeling, synthetic data, and data curation and filtering for multimodal models.
If you don't check every box above but believe you can do the job, we encourage you to apply - we're looking for capability and trajectory, not a perfect checklist match.