$224,000 - $356,500 USD yearly
Originally posted 15 September 2026 by the employer.
About the role
This role involves owning the data engine and end-to-end training verification for 3D spatial reasoning and perception capabilities within NVIDIA Cosmos models.
What you'll do
- Own the 3D data engine for Cosmos spatial reasoning, including sourcing, curating, filtering, and balancing large-scale real-world image and video corpora.
- Build annotation and auto-labeling pipelines for 3D-grounded supervision at scale, covering aspects like camera-relative 3D boxes and cross-view correspondence.
- Manage data quality end-to-end, including semantic deduplication, automated quality scoring, coverage analysis, and sampling strategies.
- Verify end-to-end model training, guard reproducibility, diagnose throughput/loss anomalies, and attribute capability changes.
- Build and operate the 3D and spatial evaluation suite, including public benchmarks such as CV-Bench, BLINK, RefSpatial, VSI-Bench, SPAR-Bench, and RoboSpatial, and NVIDIA's VANTAGE-Bench.
- Operate on large multi-node GPU clusters, tuning data throughput, sharding, and dataloader performance.
What you'll need
- MS or PhD in Computer Science, Electrical/Computer Engineering, Robotics, or a related field, or equivalent experience.
- 12+ years of experience building deep learning systems in Python with PyTorch or JAX on Linux.
- Expertise in 3D computer vision, multi-view geometry, structure-from-motion or SLAM, depth and camera pose estimation, point cloud processing, or 3D reconstruction.
- Hands-on experience with vision-language models, including building training data and evaluations.
- Experience building large-scale multimodal data pipelines: distributed video and image processing, deduplication, captioning and annotation, automated quality metrics, and dataset versioning.
- Experience running and validating large model training on multi-GPU, multi-node clusters, with knowledge of distributed training and sharding strategies such as data, tensor, and pipeline parallelism or FSDP.
- Rigorous evaluation methodology: designing benchmarks, building clean ablations, and reading results.
Skills: Deep Learning Engineer, Physical AI, 3D spatial reasoning, vision-language models, multi-node GPU clusters
This role has been open 22 days — well below the 75-day median for AI/ML Hardware Engineering roles.
AI/ML Hardware Engineering · AI ML Hardware
|
Open roles in category
521
|
Median days open
75 d
|
Median salary
$235k
|
See the full market breakdown ▾Category comparison, skills in demand, and who else is hiring
| Metric | NVIDIA | All employers we track in this specialty (521 roles · 76 employers) |
|---|---|---|
| Open roles in this specialty | 92 | 521 |
| Open roles in the wider Software, Firmware & Systems family | 964 | 6049 · 146 employers |
| Median days open | 57 d | 75 d (−18 d vs this employer) |
| Median salary (USD postings) | — | $235k |
Skills observed across this category: Deep Learning Engineer, Physical AI, 3D spatial reasoning, vision-language models, multi-node GPU clusters
Who's hiring in this category
- Qualcomm · 100 open roles · median 112 d
- NVIDIA (this employer) · 89 open roles · median 51 d
- AMD · 41 open roles · median 71 d
- Micron Technology · 30 open roles · median 60 d
- Mobileye · 19 open roles · median 70 d
- Analog Devices · 16 open roles · median 33 d
How we counted: 521 open AI/ML Hardware Engineering (AI ML Hardware) roles from 76 employers tracked in the SemiconductorJobs index, counted 8 Oct 2026. Specialty figures count only roles carrying this exact specialty label, so an employer's related work in neighbouring specialties is not included there — it is counted in the wider Software, Firmware & Systems family row. Figures refresh nightly.