CS Researcher · ML Engineer
Building at the intersection of geometric deep learning and mechanistic interpretability. Hover the field to probe the latent manifold.
ICLR 2026 @ Re-Align Research Workshop
The simplices forming above mirror the topology we exploit — sparse feature directions discovered via geometric rotation of the k-activated subspace, preserving principal components across gradient updates.
Read the PaperGradient-Preserving Sparsity · 90% dim paths · 10% principal components
Clifford Algebra · Rotor-Mediated Rotations · Bivector Orientations
The needles above are bivectors — oriented planes in Clifford algebra. Scroll to see them align via rotor transformations.
Neural architectures that respect the symmetry groups of data — SO(3), SE(3), and beyond.
Mechanistic interpretability via geometric sparsity. Finding the true basis of representation.
Models whose outputs transform predictably under input symmetries. Geometry as inductive bias.