Questions worth pursuing
Understand the system.Test its boundaries.
My research explores what model behavior reveals—and how explicit assumptions can make AI systems easier to evaluate.
Separate user and item towers produce embeddings.
OBSERVEUser & item embeddings
DISTINGUISHTwo inputs, two towers
A question with consequences.
Can the outputs of a recommender reveal whether a person’s record was used in training? ALOA studies membership inference on two-tower neural networks.
From model behavior to data boundaries.
The complementary infrastructure question is what downstream computation actually needs to receive. Elementization explores that question with an explicit capability scope and separate evaluation axes.
Explore Elementization