M
Senior Computer Vision Data Engineer
·Israel
Not specifiedFull-timeData EngineeringAutomotive
Work on a small, independent team building a unified AV dataset used to train and evaluate next-generation perception models. Focus on the curation and ML side of the data pipeline, including vision-model embeddings, scene detection, VLM-based analysis, scoring, deduplication, and sampling. Run and optimize GPU inference at scale, develop scoring strategies, and build validation tools to ensure dataset quality.
Responsibilities
- Build and improve the data curation pipeline, including vision-model embeddings, scene detection, VLM-based scene analysis, scoring, deduplication, and sampling.
- Run and optimize GPU inference at scale across thousands of driving sessions using workflow orchestration.
- Develop scoring and sampling strategies to ensure rare but important scenarios (night driving, adverse weather, hazardous situations) are well-represented.
- Collaborate with algorithm teams to identify data gaps and translate them into curation criteria.
- Build validation and diagnostics to measure dataset quality, not just pipeline health.
- Contribute to the core dataset SDK, converter, and 3D-geometry tooling (camera projection, calibration, coordinate transforms).
Requirements
- 4+ years in data engineering or backend/software engineering with production data pipelines.
- Strong Python and PyData stack (NumPy, PyArrow, Pandas, DuckDB).
- Background in research, algorithms, or ML sufficient to understand model outputs and collaborate with algorithm engineers.
- Comfort working with vision-model outputs as data (embeddings, detection results, VLM responses).
- Ability to work across team boundaries.
Nice to have
- Experience with autonomous-driving datasets or perception pipelines.
- 3D geometry and camera model intuition.
- Workflow orchestration (Argo, Airflow, Kubeflow).
- Vector databases or columnar analytics (LanceDB, DuckDB, Parquet at scale).
- Familiarity with curation concepts (active learning, hard-example mining, distribution balancing).
- Exposure to LLM agents or agentic workflows.