Background
Public evidence is limited to the Ray3 report byline, a LinkedIn listing at Luma AI, and a third-party org chart giving her title as software engineer. No papers, talks or prior employers were confirmed; a Google Scholar profile under this name was not verifiable as hers.
Luma publishes evaluation reports as a marketing and credibility artefact alongside model launches (Ray3 in September 2025, Ray3 Modify in December 2025, Ray3.14 in January 2026), so this is a recurring, productised function rather than a one-off paper.
What they run now
Evaluation of Luma's video models. The Ray3 framework she co-authored covers seven dimensions — physics and motion, instruction-following, motion artifacts, aesthetic quality, dynamic range and colour, temporal consistency, image-to-video adherence — run across thousands of generations. Luma's 2026 hiring adds an RL research role covering reward infrastructure, reward-model serving and LLM-as-judge, and a World Model role that explicitly owns metrics, which suggests the evaluation stack is being industrialised around her.
What they buy
Inference. If she owns the harness, the thing she buys is graded data and rater throughput, not tooling: calibrated human judgements delivered on a schedule, with per-rater metadata so reliability can be modelled, and ideally an API rather than a spreadsheet. Luma's stated 2026 direction toward reward models and LLM-as-judge means human labels become training data for judges, which multiplies the volume required. No vendor named.
Career
- c. 2025 – presentSoftware Engineer; co-author, Ray3 Evaluation Report, Luma AI — Title from a user-maintained org chart; not confirmed by Luma
On the record
Co-author of a framework that separates physical plausibility, instruction-following, artifacts and aesthetic quality as distinct graded axes. source