← Back to Projects & Writing

Academic Projects

Research projects and theory-first investigations spanning LLMs, fairness, generative modeling, and bio-imaging.

Research program connecting MoE routing change, residual exposure, expert correspondence, and behavioral verification
Research program · Mixture-of-Experts

MoE Routing: From Movement to Behavioral Consequence

A connected program on separating routing change from content shift, measuring residual exposure, diagnosing which expert directions receive moved gate mass, and verifying downstream behavior.

Explore the program →
Diagram of deceptive grounding: evidence retrieved for drug Y attributed to queried drug X
Paper

Accurate, Grounded, and Wrong

Deceptive Grounding: a clinical RAG system relays a real drug's evidence faithfully but staples it onto the wrong drug, slipping past every standard hallucination, faithfulness, and citation check. 7.8% in production, and worse for medically fine-tuned models.

Clinical RAG LLM Safety Evaluation
Microscopy image of glowing cells
Paper

Foundation Model for Cell Painting (CellPainTR)

A novel contrastive learning framework to explicitly disentangle biological signals from nuisance variables (like batch effects) in high-content imaging data.

Transformers Self-Supervised Bio-informatics
Academic Project

Mitigating Racial Bias in Face Age Progression

An investigation into fairness for generative models. We modified a StyleGAN-based age progression model to mitigate racial bias by promoting more equitable transformations across demographics.

Generative Models Fairness in AI StyleGAN