Within Perception, the Data Refinery team builds the data and evaluation systems to benchmark models and accelerate the ML development cycle. We design the centralized systems and tools for rigorous model quality analysis, debugging, and performance measurement after a model has been trained. Our work provides the fast, actionable feedback loops that shorten the development cycle from days to hours , enabling teams to iterate, improve, and ship better models faster. We collaborate across ML and Infrastructure teams to integrate this standardized evaluation into the core ML lifecycle and help bring new capabilities to the Waymo Driver.

In this hybrid role, you will report to an engineering manager.

You will:

• Design, build, and maintain scalable data pipelines to process many petabytes of complex sensor data, making it ready for efficient model training and evaluation.
• Develop infrastructure to produce reliable, high-quality datasets for a wide range of ML models, from real-time on-car models to large-scale offboard foundation models.
• Build and scale a centralized model benchmarking system to provide fast, standardized, and reproducible model quality metrics, reducing the model development cycle from days to hours.
• Implement tools for deep model quality analysis, including comparative leaderboards and integrated triage/debugging workflows.
• Automate data quality and validation checks to ensure the integrity, consistency, and trustworthiness of our datasets as we scale to new cities and vehicle platforms
• Collaborate closely with ML engineers, research scientists, and core infrastructure teams to understand user needs and deliver impactful data and evaluation solutions.

You have:

• C++ programming skills (required), with Python experience

• 5+ years in ML data engineering, including data pipelines, data curation, data balancing.

• Experience with the ML development lifecycle, including data engineering, model evaluation, and metric computation.

We prefer:

• Hands-on experience in building large scale data processing or retrieval systems and pipelines: Apache Spark, Apache Beam, Google Cloud Dataflow, AWS Data Pipeline, Faiss/ScaNN, etc.
• Experience with the end-to-end process of benchmark creation, from defining domain-specific metrics, to curating challenging evaluation datasets, to triaging examples and improving the benchmark.