Strong machine learning fundamentals with proficiency in Python and PyTorch.
Experience in training large vision models from scratch, not just fine-tuning existing ones.
Depth in a computer vision specialty like pose estimation, SLAM, or 3D reconstruction.
Proven track record of owning models in production and maintaining responsibility for them.
Ability to work with noisy real-world data and adapt to shifting priorities.
Responsibilities
Build models to address current challenges using large multi-view, multi-sensor datasets.
Tackle complex footage issues such as motion blur and occlusion.
Evaluate model performance against real-world video and benchmark results.
Optimize GPU performance and ensure reproducibility in production environments.
Investigate and resolve customer-reported issues, delivering effective fixes.
Benefits
Collaborative work environment focused on innovation.
Opportunity to work on cutting-edge technology in computer vision.
Encouragement for continuous learning and professional growth.
Flexible work arrangements to support work-life balance.
Full Job Description
The role
Mecka captures how people move, work and interact with the physical world. The footage is messy: hands disappear behind objects, cameras turn quickly, and motion blur hides the details. You'll build models that make sense of it-and turn it into accurate 3D data that robotics teams can train on.
We're looking for machine learning engineers with depth in pose estimation, SLAM or 3D reconstruction who want to own what happens after the experiment works. You'll train models, investigate where they fail and get them running reliably in production, working across the perception stack along the way. What you'll do
Build models for the problems in front of us. You'll train vision models from scratch on large multi-view, multi-sensor datasets, developing new losses and architectures where existing methods fall short.
Work through the difficult footage. You'll tackle motion blur, fast rotation, heavy occlusion and partial views of hands, bodies and objects.
Find out where your models fail. You'll evaluate them on our own video, measuring real error and failure rates alongside public benchmark results.
Own the path to production. You'll optimize GPU performance, make results reproducible, integrate models into our data pipeline and ship releases you can roll back.
Follow problems through to a fix. You'll reproduce customer-reported issues, understand what went wrong and deliver a fix. You'll also prototype quickly as new perception problems come up.
What you'll bring
Strong machine learning fundamentals. You're comfortable building and training models with Python and PyTorch.
Experience training large vision models from scratch. Your experience goes beyond running or fine-tuning existing models.
Depth in a computer vision specialty. You've worked deeply in at least one area, such as pose estimation, SLAM or 3D reconstruction.
Experience owning models in production. You've shipped a model or perception component and stayed responsible for it afterward.
Comfort with the work as it is. You can make progress with noisy real-world data and adapt when priorities shift.
Nice to have
Experience in one or more of these areas would be valuable:
Hand and body tracking: 3D hand and full-body tracking, mesh recovery, human-object and human-scene interaction, and estimating contact with objects and the ground.
SLAM, VIO and calibration: Learned ego-motion, optical flow, online calibration and metric scale.
3D reconstruction and objects: Feed-forward or per-scene reconstruction, Gaussian splatting or NeRF, and neural 3D object estimation from egocentric video.
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.