Research Engineer, Preference Data

Vizcom•San Francisco, CA

1d•$250,000 - $400,000•Onsite

About The Position

Vizcom is a Series B company that provides a platform for design teams at major companies like Nike, GM, New Balance, and Hasbro to perform daily design tasks including sketching, rendering, color and material selection, 3D modeling, and exporting. The platform generates a significant amount of data from over 700,000 designers, which is considered a valuable asset. Currently, this data is more like historical records than a usable asset, with only a small fraction being suitable for training purposes. The role of a Research Engineer is to transform this raw data into a machine-usable format for training models. A key challenge is that design sessions are branching trees, not linear sequences, and the current data capture only records single steps, missing the nuances of design choices like forking, backtracking, and abandonment. As an ML data engineer, you will be responsible for building the system that converts professional design work into training-grade preference data and uses training results to improve the product. You will collaborate closely with researchers who use the data you build, work within the product code where data signals originate, and interact with the data warehouse where the data lands. The role is highly interactive, with direct feedback from users within days of shipping datasets. This is not a support or offline ETL role; the pipelines you create will operate within a live canvas used by professional teams under enterprise agreements with major brands. The core challenge is to capture more data without compromising trust or performance. The ideal candidate believes that future advancements will be driven by data and is motivated to build the systems that feed machine learning models.

Requirements

Have built training-data or large-scale data pipelines that other people depended on.
Have an experimentalist mindset.

Nice To Haves

Have trained models yourself, so you know what training actually consumes.
Have run labeling, human-feedback, or evaluation ops.
Have worked with data under real privacy or contractual constraints and enjoyed the puzzle.
Look at product exhaust and see evidence.

Responsibilities

Build the system that turns professional design work into training-grade preference data.
Build the system that uses training results to improve the product.
Work beside researchers consuming the data.
Work within the product code where data signals originate.
Work against the data warehouse where data lands.
Design and build the capture surface in collaboration with product engineers, shipping it in product code.
Develop the pipeline from canvas to warehouse to training set, ensuring it is clean, versioned, and reproducible from a fingerprint.
Establish dataset contracts and lineage, ensuring every example is traceable to its origin and every result is replayable months later.
Build the boundary system to manage data capture under enterprise agreements, including consent, isolation, and sanitized slices for research and work trials.
Develop collection instruments to gather explicit feedback from designers when historical signals are insufficient.
Ensure the honest encoding of ambiguity in the data, distinguishing between unpicked, disliked, abandoned, and rejected actions.