Data Pipeline Manager

Role Overview

Before a good model, you need good data and a good evaluation framework.

WoRV's Data & Evaluation Manager is not someone who accumulates data — they are someone who designs and owns the entire data pipeline. This role defines what data the product needs to improve, what criteria determine that something is "ready to ship," and then operates the full collection-refinement-evaluation loop against those criteria.

WoRV continuously searches for the best methods to acquire high-quality data quickly. That could mean running a GPS-based semi-autonomous fleet in the field to collect real-world data, or leveraging operational data from VLA-based autonomous driving. The collection method is not fixed. The essence of this role is defining and accelerating the first cycle of the data flywheel — bridging the gap between the data the Research team (VLA/VLM) needs and the data that can realistically be collected from customer sites.

WoRV is currently collecting data across multiple industry projects simultaneously — agriculture, construction, and ports among others. For data from individual projects to accumulate into shared model capability for the entire team, standardized formats, metadata schemas, quality standards, and collection efficiency management are required. The Data & Evaluation Manager owns all of this.

Responsibilities

1. End-to-End Data Pipeline Ownership

• Define field data collection strategy and maximize collection efficiency (ratio of useful data to setup time).
• Design and operate the end-to-end pipeline for sensor, teleoperation, and field operation data.
• Define data schemas, versioning, metadata frameworks, and standard formats (LeRobot, etc.).
• Manage storage infrastructure (DGX SSD, NAS HDD, etc.) and the training data lifecycle.

2. Evaluation Framework Design and Operations

• Build an evaluation framework that distinguishes customer-specific KPIs from shared product KPIs.
• Connect model-, module-, and system-level offline and online evaluation.
• Build an experiment framework that quantitatively demonstrates the "data → model performance" relationship.
• Surface failure cases and design retraining / re-evaluation loops.

3. Cross-Project Data Strategy

• Coordinate data collection priorities across multiple projects (Navigation, Manipulation).
• Collaborate with the Research team to define collection scenarios and topics.
• Build structures that allow individual project data to accumulate into shared model capability.
• Work with PM to design customer contract structures around data ownership and usage rights.

Qualifications

• Experience designing and operating data pipelines or data platforms
• Experience formulating data collection strategies and connecting them through to field operations
• Ability to judge "what data we need" from a problem-first perspective
• Ability to create metrics and design them so they don't pull the team in the wrong direction
• Ability to communicate with and coordinate priorities across multiple stakeholders: Research, PM, and field operations
• Data processing skills in Python, SQL, etc.

Preferred Qualifications

• Experience building and operating CV / Robotics / Autonomous Driving datasets
• Experience with labeling ops, taxonomy design, and annotation quality control
• Experience with evaluation harnesses, benchmarks, and experiment tracking
• Experience processing multi-modal sensor data (camera, LiDAR, IMU, etc.)
• Experience with active learning, hard case mining, or data flywheels
• Experience with the HuggingFace ecosystem (LeRobot, Datasets, etc.)
• Experience with MLOps or data infrastructure
• Experience leading or managing a data team

Who We're Looking For

• Someone who treats data design as more important than model training
• Someone who relentlessly digs into "why did we fail because this data was missing?"
• Someone who can see the full flow from collection site to training server and find the bottleneck
• Someone who views data as an asset and designs for reusability