Generative and geometric machine learning — developed rigorously, deployed at scale, for scientific discovery.
The boundary between mathematics, physics, and machine learning
is dissolving. Transformers, diffusion maps, and stochastic dynamics
are not separate tools — they are different regimes of the same geometry.
DAC is built on this unification. We develop the infrastructure
to run these methods at scales that matter to science.
Coifman–Lafon diffusion maps at 100M-sample scale. B300-native sparse eigensolver. Hours, not weeks. Built for MD trajectories, cryo-EM, XFEL, and beyond.
Additional solutions in development.
Documentation is being written in parallel with the codebase. API reference, tutorials, and theory-to-code guides will ship here as the platform moves out of private beta.
Submit your first run. Dataset formats, parameter reference, and reading your results.
In progressEndpoint schema, field definitions, response format, and error codes for the FlashDiffusion API.
In progressHow Coifman–Lafon diffusion maps translate to the FlashDiffusion pipeline, step by step.
Coming soon// In the meantime: arXiv 2604.09560 is the primary reference · GitHub has the source
View documentation →The Diffusion–Attention Connection
Transformers, diffusion maps, and magnetic Laplacians are different regimes of a single Markov geometry built from query–key scores — unified through the static Schrödinger bridge framework. DMAP is the equilibrium sector. Attention is NESS. The same equation governs both.
Read the theory →Row-stochastic transition matrix. Encodes manifold geometry, not sampling density.
F = 0 gives DMAP (equilibrium). F ≠ 0 gives attention (NESS).
Eigenvector embedding weighted by eigenvalues. Euclidean distance = diffusion distance.