Here’s a list and short description of some of my past projects.

Royr: generalizable imitation learning with robot memory and object-based priors. (In Progress)

We aim to use object-based methods to improve imitation learning to address challenges in learning from real-world data such as different camera viewpoints, object task relevance, background changes, and occlusion. My contribution so far has focused on implementing an object-oriented diffusion policy pipeline, utilizing embeddings from vision foundation models (Dino V2, CLIP, Theia) and Pointnet++ for point clouds, RPN/Segment Anything, and object-wise attention/transformer

Validating the Usage of Stable Diffusion Foundational Models on Generating Glaucoma Fundus Images in Low-Data Settings

Final project for my Computer Vision class which focused on generating synthetic image data in low-resource settings for medical machine learning to better protect patient privacy and improve patient diagnoses. I utilized stable diffusion foundational models and Generative Adversarial Networks to perform this task. My methods improved Kernel Inception Distance by 53% from baseline using fine-tuning and prompting and improved model recall on glaucoma diagnosis on the JustRAIGs dataset by 2.5%.

2024 Citadel Women’s Datathon Report - 1st Place Winner

I won first place at the Citadel Women’s Datathon for my team’s analysis on regional differences in the impact of gas prices depending on oil production and transit ridership. Our analysis utilized various data visualizations with Matplotlib and statistical modeling/testing to validate our conclusion.

AutoManim - TreeHacks 2024

My project for TreeHacks, which sought to generate Manim animations (used by 3 Blue 1 Brown) given a user prompt. I used techniques inspired by Grammar Prompting for Domain-Specific Language Generation with Large Language Models by Wang et al. as well as Large Language Models are Zero-Shot Reasoners by Kojima et al. to develop a domain-specific prompt utilizing Backus Naur Form, which allowed us to obtain high performance with a total token usage 50 cents over the course of the entire hackathon. I helped create our website with the MERN stack.

RoomCraft - Web.lab Honorable Mention

I placed in the top 7 teams out of 450+ MIT participants at the Web.lab programming class/competition for my team’s website, which seeks to seamlessly combine productivity with gamification. Our website was made with React.js/HTML/CSS for the frontend, Node.js/Express/MongoDB for the backend, and ReactQuill/OpenAI API.

BioAcoustix - HackMIT

Our project was a platform providing bioacoustic indicators for telehealth to predict diseases from user audio recorded from a web platform. I used transfer learning from VGG16, melspectograms, Wav2Vec, HuggingFace, Librosa, Keras, Tensorflow, and sk-learn to train the AI models used at our website’s backend.

Systematic Optimization of App Generation Few-shot Learning for LLMs Trained on Code - Regeneron STS Scholar

My project for Regeneron STS that placed in the top 300 in the US, which focused on the synthesis of few-shot prompts for mobile app generation. I developed a novel algorithm called p-mRMR which adapts the mRMR feature selection technique to prompts by subbing correlation with cosine similarity, which improved performance by 43% for “simple” app generation tasks and 55% for “complex” app generation tasks.

Utilizing Explainable Machine Learning for Determining the Etiology of Unexplained Dyspnea

I created an explainable visualization platform for the diagnosis and etiology of dyspnea (shortness-of-breath). I utilized various feature selection techniques, data processing/imputation, bayesian statistics, and various ML models using sci-kit learn. My research found significant differences in the risk factors/models for patients of differential gender, age, and BMI.