Portfolio

CS224R final project investigating whether reinforcement learning with verifiable rewards (RLVR) can expand a small language model's reasoning frontier beyond its pretrained action space. Tests tool-integrated RLOO against the "invisible leash" hypothesis using Pass@K evaluations.

• Reinforcement Learning
• Language Models
• RLVR
• Reasoning
• Tool Use

Survey paper examining sensor-based and AI-enabled approaches to olfactory assessment, including electronic-nose chemical sensing, VOC analysis, wearable nasal airflow monitoring, and olfactory-evoked EEG responses, framing olfaction as both a diagnostic biomarker and an intervention target for early detection of neurodegenerative decline.

• Medical AI
• Neurodegeneration
• Sensor Technology
• Early Detection
• Biomarkers

Developed deep learning models for automated brain tumor segmentation using the BraTS 2021 dataset, comparing CNNs, transformers, and promptable architectures for pixel-level tumor detection.

• Medical AI
• Computer Vision
• Deep Learning
• Semantic Segmentation
• Healthcare Technology

Built a custom CNN with adversarial training to detect Coca-Cola logos under blur, noise, and occlusion. Achieved +13% accuracy over YOLOv8 with extensive data augmentation.

• Computer Vision
• Adversarial Robustness
• Object Detection
• Data Augmentation

Implemented a full implicit + explicit free list allocator in C, including malloc, free, and realloc. Built debugging utilities (validate_heap, dump_heap) and stress-tested on real allocation traces.

• Systems Programming
• Memory Management
• C
• Performance Optimization

Trained VGG-style CNNs, GRUs, and LSTMs on mel spectrograms from GTZAN, incorporating noise, pitch-shift, and time-stretch augmentations for robustness.

• Audio Classification
• Deep Learning
• CNNs
• Music Information Retrieval

Built a GPT-4–powered browser extension that detects persuasive fallacies in political news, provides real-time annotations, and generates "extremeness" scores. Ran a pilot RCT to evaluate behavioral impacts.

• Human-AI Interaction
• Politics & Psychology
• Media Literacy
• Language Models

Benchmarking Grad-CAM, Integrated Gradients, and GradientSHAP across segmentation models, combining quantitative evaluation with clinician + researcher feedback to assess clinical usability.

• Explainable AI
• Medical Imaging
• Model Interpretability
• Human-AI Interaction

Co-authored research manuscript with Stanford School of Medicine and the University of Utah. Retrospective cohort study of 2,814 graduate medical trainees (2014–2020) analyzing gender and racial disparities in ACGME Milestone progression across 107 hospital-based, medical, and surgical programs using Wilcoxon rank-sum, Kruskal-Wallis, and linear mixed-effects models.

• Medical Education
• Health Equity
• Biostatistics
• Bias in Assessment
• Longitudinal Analysis

Examines how to govern frontier AI before catastrophic risks materialize, proposing risk-monitoring taxonomies, cross-lab oversight architectures, and pathways from voluntary commitments to binding regulation.

• AI Governance
• AI Safety
• Policy
• Institutional Design
• Existential Risk