Massachusetts Institute of Technology

B.S. in Computer Science
Selected Coursework
- Graduate Machine Learning, Computer Vision, Design and Analysis of Algorithms, Fundamentals of Statistics

Hudson River Trading | Women in Trading and Technology Intern

Investigating Human-LLM communication and LLM representations of concepts like cats or ballerinas through the medium of sketches
Moving LLM inference pipeline to Llama 90B run on GPU cluster
Using quantization, performance optimization, HuggingFace accelerate library

Best Paper Award @ CoRL LangRob Workshop, under review at ICRA
Programming novel keypoint-conditioned diffusion policy; evaluating 20+ policy network variants utilizing Linux, PyTorch, HPC to improve robotic generalization; integrating multimodal models utilizing pointcloud encoders, object-wise attention/transformers, and vision foundation models, improving evaluation performance by 27.5%
Creating Visual QA keypoint proposal pipeline with SAM segmentation, furthest point sampling, mark-based prompting, nonlinear optimization/rejection sampling (RANSAC)

Trained Python machine learning models predicting dermatology product sentiment from 3000+ amazon reviews and ingredient data, narrowing 6680 features to 10 most relevant features using feature selection and improving Random Forest/XGBoost performance by 6.55%.
Developed an automated web scraping framework for the PubChem site, facilitating the retrieval of data on 1.11 million compounds; enhanced efficiency by improving ingredient-compound association accuracy by 30%.

Using Relation-Aware Graph Attention Networks to create commonsense-informed language embeddings

Conducted at Research Science Institute, received Regeneron STS Scholar(formerly Westinghouse)
Optimized pipeline, evaluated 10800 lines of generated code for testing few-shot prompt synthesis for LLMs
Designed a novel few-shot selection algorithm p-mRMR, improving application generation performance by 55%

Built an optimized compiler in C++ for the Brainf*ck coding language, achieving 1000x performance speedup
Utilized intermediate representations, jump tables, pragmas, bitwise compression to enable efficient execution
Won HackMIT challenge for our fast implementation, receiving a prize of $1000.

Created full-stack resume website in Typescript with React.js/HTML/CSS frontend, Node.js/Python/AWS backend.
Utilized a Serverless Application Model with Cloudfront, REST APIs, CloudFormation, Lambda, and S3 hosting.
Built automatic CI/CD pipeline using Github Actions with pytest and Playwright for unit testing/integration testing.

Authored 10-page data science report in 8 hours about the regional impact of gas price flunctuations on poverty.
Collaborated with teammates to utilize/clean four external datasets, perform the Augmented Dickey-Fuller test for time series nonstationarity and Granger Causality - Testing for causal inference, and train four regional VAR models.
Won first place out of 80+ invitees from 2000+ applications across North America, winning a prize of $10,000

Researched the generation of synthetic medical images using stable diffusion foundation models and Generative Adversarial Networks on glaucoma images, improving Kernel Inception Distance by 53% from baseline and deep residual neural network (Resnet) recall on glaucoma diagnosis on the JustRAIGs dataset by 2.5%.

Developed a website combining productivity with gamification, with a React.js/HTML/CSS frontend, Node.js/Express/MongoDB backend, and ReactQuill/OpenAI API. Utilized Figma to generate mockups.
Utilized Object-Oriented Design in file/note-taking system, user profile/friending, and game-like user interface
Placed in the top 7 teams out of 450+ MIT participants, earning an honorable mention and $750.