My Projects
WattBot: Estimating AI Emissions with Retrieval Augmented Generation (RAG) (2025)
As part of the UW-Madison Machine Learning Marathon, I collaborated with a team of five to design a retrieval-augmented generation (RAG) system that measures and interprets the environmental impact of AI models. The goal of this project was to create a tool that could extract and summarize peer-reviewed data on the carbon and water footprints of AI systems, providing transparent, citation-backed insights for researchers and policymakers.
We built WattBot using retrieval-augmented generation (RAG) pipelines, vector-based document retrieval, and cloud-hosted inference infrastructure to analyze AI sustainability research and generate grounded responses. The system combined document parsing, semantic search, and source-aware generation to ensure outputs were traceable to verifiable evidence, enabling more transparent and trustworthy insights at the intersection of AI, energy usage, and policy.
Through this project, we demonstrated how RAG systems can promote accountability and sustainability in AI by making complex environmental data more accessible and actionable.
Virtual Dentist: Using Large Language Models to Simplify Doctor-Patient Interactions (2024)
For my senior project, I partnered with a local dental clinic to simplify patient–doctor interactions by building a chatbot that recommends treatments based on patients’ stated problems and medical history. The core idea is to reduce the initial communication gap between patients and doctors by providing accessible, personalized oral-care guidance before an in-person consultation.
I designed a dental diagnosis chatbot that extends ChatGPT’s capabilities using large language models and Python. The chatbot takes as input a structured dataset containing patient history, medical records, reported symptoms, and training data. Based on this information, it can recommend oral-care tips, suggest over-the-counter (OTC) medication for pain relief, or even provide remedies if the patient has experienced similar issues in the past.
To support the project, I created a mockup website in Figma and developed a sample database of over 100 patient records in collaboration with my external advisor and local dental professionals. This database strengthens the chatbot’s ability to provide more accurate and context-aware diagnoses by referencing historical cases.
Finally, I conducted model optimization using diagnostic evaluation metrics and achieved an accuracy rate of 85%. By combining domain expertise with large language models, this project demonstrates how AI can assist both patients and doctors in improving the efficiency and accessibility of dental care.
Volleyball Activity Recognition Feedback Mechanism (2025)
In 2025, I worked on a project with Inspirit AI focused on volleyball activity recognition. The goal of this project was to develop an algorithm that could classify player actions and predict positions on the court based on movement and location, ultimately providing insights to improve athletic performance.
I built the model using Python with Keras and NumPy, incorporating OpenCV for video preprocessing and a convolutional neural network (CNN) for action classification. To strengthen the dataset, I performed data preprocessing by filtering out irrelevant images and ensuring that only volleyball players performing specific tasks were included. This refinement improved both the accuracy and relevance of the model’s predictions.
As a final component, I developed a feedback system that provided players with tailored suggestions on how to improve their performance. By combining computer vision with machine learning, this project demonstrated the potential of AI-driven systems to analyze sports performance in real time and deliver actionable insights to athletes.
Music Recommending Chatbot (2022)
In 2022, I developed a music-recommending chatbot designed to suggest songs and generate automated responses based on a user’s mood for the day. Built using Python and CMU Graphics, the chatbot aimed to create a more personalized and engaging music discovery experience by tailoring its recommendations to user input.
To power the system, I integrated a database containing 20+ diverse music genres and tracks. I then refined the application by analyzing user feedback, which allowed me to improve both the accuracy of the music recommendations and the responsiveness of the chatbot’s interactions. This project highlighted how AI can blend user sentiment analysis with entertainment, creating a more customized listening experience.