Projects

Extending Tensor Depth Efficiency to Tensor-Train and RNN-like Models

This project extended a classic tensor-analysis paper on deep learning expressivity by replacing the original CP/HT setup with a tensor-train decomposition, connecting the result to an RNN-like architecture and deriving an analogous exponential rank lower bound.

Theory Tensor Decomposition RNNs

Inducing and Analyzing Hallucination in Diffusion LLMs
Pinned

We studied how factual hallucinations can be induced in diffusion-based language models—which iteratively denoise entire sequences rather than generate tokens autoregressively—to understand what these failures reveal about their internal behavior.

Hallucination Diffusion LLM

Exploring Beyond VisRAG
Pinned

We evaluated multimodal RAG on visually simple but text-heavy documents to analyze where VisRAG breaks down, how far text-only retrieval can go, and whether combining the two modalities produces a stronger retrieval system.

RAG VLM Multimodal

Multilinear Operators for Arithmetic Tasks
Pinned

This project explored whether sequence-to-sequence models could be built entirely from multilinear operations, without softmax attention, nonlinear activations, or positional encodings, and resulted in MuSe, a multilinear sequence model designed for arithmetic tasks such as addition, multiplication, and parity.

Polynomial networks Tensor operators

Gaze-aware Proactive AI Assistance for Digital Article Readers

This team project explored a browser-based reading assistant for ESL readers that combines webcam gaze tracking, document-level gaze localization, and LLM-based paragraph explanations to reduce friction when reading difficult articles.

HCI Eye Tracking LLM Systems

PseudoPremium: Optimizing Visitor Experience at Universal's Islands of Adventure

This project formulated a mixed-integer optimization model for theme-park itinerary planning, balancing visitor preferences, ride wait times, ride durations, express-pass decisions, and seasonal crowd conditions to maximize overall satisfaction.

Optimization MILP Operations Research

Juno Meets Jupiter: Recreating and Reimagining JunoCam Imagery

This Mathematica project combines parameter search, channel-wise reconstruction, and scientific visualization to recreate and reinterpret raw JunoCam imagery.

Image Processing Scientific Visualization Mathematica

Something Fishy This Way Comes: Fish Segmentation in Mathematica

This project shows how I built and debugged an interpretable fish-segmentation and measurement pipeline in Mathematica using color-space analysis, morphology, and connected-component filtering.