Five years shipping AI from research to production.
Currently building AURA — a 30B multimodal foundation model generating 2K video + synchronized stereo audio in a single forward pass, with sampling cut from 50 steps to ≤8 via Rectified Flow + step distillation.
"A model that only works at fifty sampling steps is a research result, not a product."
The through-line across all of it: inference budgets — getting capable models to run fast, cheap, and reliably outside the lab.
Correctness fixes over cosmetic changes — every fix ships with a failing-then-passing test, and each one is reproduced and root-caused before a line is written.
8 merged upstream · 11 more in review
Generative AI & ML
Languages & Frameworks
Cloud & MLOps



