Own the path from raw capture to training-grade data, and the embodied models we fine-tune on it.
The role
We record first-person footage of skilled physical work with our own multi-camera capture rigs, and we fine-tune embodied models on what we capture. You own everything between the two: the pipeline that turns raw multi-sensor recordings into training-grade datasets, and the fine-tuning runs that prove what the data is worth.
A training run is only as good as its worst batch. Your quality gates decide what gets in.
What you'll do
- Build the pipeline from raw capture sessions to training-ready datasets: ingest, sync validation, deduplication, QA gates.
- Fine-tune embodied and vision models on our captures, and iterate from failure analysis.
- Define data-quality metrics and validation gates. Root-cause regressions across the pipeline.
- Build auto-labeling and human-in-the-loop annotation tooling that scales past manual review.
- Work with multi-sensor streams: multi-camera video, IMU, timestamps that have to agree.
- Translate model failures into concrete changes to what and how we capture, working with hardware and field operations.
- Run and track training jobs: evals, dashboards, regressions caught before they cost a week.
What we look for
- Hands-on experience training and fine-tuning vision or embodied models, shipped into production.
- You've built large-scale data pipelines and have a quantitative mindset about data quality.
- Strong Python and PyTorch. You own problems from raw bytes to training curves without waiting for permission.
- You debug datasets with the same rigor as models: most failures live in the data.
Nice to have
- Imitation learning or VLA fine-tuning: ACT, diffusion policies, or similar.
- Egocentric video understanding and datasets like Ego4D or Epic-Kitchens.
- Multi-view geometry, SLAM, or hand-pose estimation.
- GPU cluster training infrastructure, and model optimization for deployment.
#J-18808-Ljbffr
