ML Theory
Representation Learning
How models learn structured, reusable representations of raw signal — and what that structure reveals about the data itself.
EmbeddingsProbingManifold structureContrastive objectives
A throughline across every project: whether the input is a conversation, a collider event, or a stream of sensor data, the interesting question is always what geometry the model discovers, and whether that geometry matches the structure we already know to be true. This work focuses on probing, visualizing, and shaping learned embedding spaces so they stay interpretable as they scale.