I lead product for data & model capabilities at Luma AI, working on Uni-1, Ray 3, and the Keyframes feature. Currently thinking about how to turn a research breakthrough into a feature.
Luma's first unified understanding & generation model — a decoder-only autoregressive transformer where text and image live in one interleaved sequence. Plans composition before rendering it.
Studio-grade 16-bit HDR, character reference, visual annotation, and a reasoning loop that lets the model critique its own output before committing. Shipped Sept 2025 — Ray3 Modify followed in December.
Pick a start image, pick an end image, prompt the middle — the model authors the transition. We turned video generation from a one-shot prompt into a controllable narrative primitive.
An animated map of how Luma's models have evolved across capabilities — turning abstract benchmark progress into something you can feel.
Check it outWeekend experiments that test product ideas at the speed of frontier models.
Graph-native UI for navigating Claude's reasoning — treats an agent run as a structure you can branch, prune, and replay instead of a flat chat log.
Try it
A scripted pipeline that spins prompts into short-form vertical drafts using frontier video models — built to feel native to how creators actually iterate.
Try itI gather people. The three things that have shaped me most outside of work:
I host the occasional gathering in San Francisco — find past ones on my Luma profile. If any of the three above are your thing, say hi.