What will your job look like: Work collaboratively with shared ownership. Your focus area will be the curation and ML side of our data pipeline, but you will contribute across the full stack alongside the rest of the team. Build and improve the curation pipeline - from vision-model embeddings and scene detection, through VLM-based scene analysis, to scoring, deduplication, and sampling that produces a balanced and diverse dataset. Run and optimize GPU inference at scale (embedding extraction, VLM inference) across thousands of driving sessions using workflow orchestration. Develop scoring and sampling strategies that ensure rare but important scenarios (night driving, adverse weather, hazardous situations) are well-represented in the final dataset. Work with algorithm teams to understand what data gaps hurt model performance and translate those into curation criteria. Build validation and diagnostics that measure dataset quality - not just pipeline health, but whether the data is actually good for training. Contribute to the core dataset SDK, converter, and 3D-geometry tooling (camera projection, calibration, coordinate transforms). All you need is: 4+ years in data engineering or backend/software engineering with serious data work — pipelines that run in production, not just notebooks. Strong Python and the PyData stack (NumPy, PyArrow, Pandas, DuckDB). Some background in research, algorithms, or ML — enough that you can read a paper, understand a model's outputs, and have informed conversations with algorithm engineers. Comfort working with vision-model outputs as data: embeddings, detection results, VLM responses. Ability to work across team boundaries — this role lives between algorithm teams, infra teams, and our own. Advanteges: Experience with autonomous-driving datasets or perception pipelines. 3D geometry and camera model intuition (or the mathematical background to ramp up). 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) — useful context, not a requirement. Exposure to LLM agents or agentic workflows for data tasks.
