Infrastructure for geometric and generative machine learning at the scales that matter to real scientific datasets.
Coifman–Lafon diffusion maps at 100M-sample scale. A B300-native sparse eigensolver replaces the dense EVD bottleneck — running in hours what previously took weeks. The geometry of your data, at the scale of your data.
Diffusion Image Transformer Operator. A flow-based DiT that trains on your image dataset — pixel-space, class-conditioned, multi-GPU ready. Submit a job, configure your architecture, and checkpoints ship to R2 as training runs.
Running large-scale MD, cryo-EM,
or XFEL data? Let's talk.