Papers & Research

Work that leans more into research, evaluation, and writing than product UI, but still grounded in concrete systems and data.

Benchmarking Commonsense Visual Reasoning for Vision-Language Models

Co-author with Prof. Ernest Davis (NYU) · arXiv cs.AI · Submitted Mar 2026

Visibility-reasoning benchmark for vision-language models with a 2×2 XOR design across 100 families and 300 headline cells, with automated graders and scoring.

My role

Evaluated 9 VLMs across three tiers (flagship, prior-gen, open-source)
Surfaced a 26% abstention rate in GPT-5 as a key finding
Top scores: GPT-4o (0.728) and Gemini 3.1 Pro (0.727) effectively tied

View code/project

Adversarial Robustness of 3D U-Net Radiation Therapy Dose Prediction

Co-author with Prof. Birjoo Vaishnav (U. Maryland) & Prof. Weixian Liao (Towson) · Accepted at SERA 2026 (IEEE/ACIS); submitted to ASTRO 2026 · 2025 – 2026

Evaluated adversarial (FGSM/PGD) and clinically plausible CT perturbations on the OpenKBP dataset using clinical DVH metrics.

My role

Identified resolution downsampling as the only perturbation with immediate clinical impact
Bone shift late takeoff at level 4-5; noise, bias field, and dental flat through extreme amplitudes
Junior Investigator Travel Award application for ASTRO 2026

View code/project

Legal Text Classification with BERT and LegalBERT

Co-author · NYU Natural Language Processing · Course project · 2024

Fine-tuned BERT-family models on legal corpora; compared against classical baselines with reproducible training and reporting.

My role

Compared BERT-Double, Legal-BERT, and Custom-Legal BERT models
Implemented end-to-end NLP pipeline over legal case corpora
Proposed and implemented a composite evaluation metric for legal QA

View code/project Read paper (PDF)