In this project, I aimed to create a fun, lightweight experience for users to explore and enjoy. I chose to focus on generative artwork, demonstrating how even a small amount of training data, combined with a pre-trained model originally designed for a different purpose can produce creative and unexpected results. This project highlights the flexibility and potential of adapting AI models for playful, interactive applications.

Generative AI is a fast-growing niche in the world of machine learning with applications in various industries, including AI generated art. To expand on this niche, my team has fine-tuned an image diffusion model to generate pixel-style GIFs that morph between two images from provided text prompts. This involved fine-tuning an existing diffusion model on pixel-art dataset(s), generating two images from the first and second prompt, and interpolating their prompts and latent noise tensors together to morph the generated images. To create pixel-art GIFs we extracted the interpolation outputs and stitched the outputs together. Our work aims to narrow down the use cases of diffusion models and contribute to the field of generative modeling.

In this project, I attempted to use unsupervised machine learning algorithms to analyze SEC 10-K filings. The goal was to see if we could find any hidden insights by only looking into companies' publicly available financial statements. As expected, more data is necessary to fully understand any given company, but it was still interesting to apply some simple algorithms to this data.

Companies are required by law to file the 10-K form annually. This includes an overview of the firm's business, risk factors, selected financial data, management's discussion and analysis, and financial statements. We have analyzed a snapshot of this information from 2020 using unsupervised learning algorithms and discovered some hidden insights. Mainly, we have attempted to cluster the companies based on the text and tabular financial data. We have found that some clusters did reflect some intuitive real-world meaning after we performed thoughtful preprocessing and algorithm selection. However, we still didn't capture the full depth of structure that we expected, suggesting that more data would be needed to truly understand a company.

I developed this project in my spare time as a fun way to combine my hobby of table tennis with technology. The goal was to automatically track scores during a game. Currently, the system uses a mobile phone camera for tracking, but it could be enhanced with a professional-grade camera with a higher frame rate for improved accuracy and smoother performance.

This project tracks a table tennis ball as it moves around during gameplay. While in professional table tennis, the ball may be hit at much faster speeds, this project aims to track the ball at lower speeds using a regular mobile device camera. Additionally, the performance is highly dependent on the lighting conditions and overall environment setup. There are two ball tracking algorithms that are being used, both of which use a color based image map for orange table tennis balls. The first algorithm detects the contours of orange blobs. The second algorithm uses hotbox detection.

In this project, I developed a complete hardware + software package for users to play the game of Pacman and showcased it at Cornell Tech's open studio. Our project was also featured on Cornell Tech's social media: LinkedIn, Instagram.

Our project is a handheld matrix gaming console in which we have developed the game of Pacman from scratch. By tilting the board, the user can control the position of Pacman and eat food or ghosts to accumulate points. There's a catch though! The louder the environment the user is playing in, the more difficult the game. So put your skills to the test and play a game of Pacman!

My team participated in the HackGPT 2023 hackathon with a mission to help researchers navigate academic archives with our tool called InstantResearch.AI. Utilizing techniques like semantic embeddings and prompting, we've developed a system that streamlines the search for relevant research papers. Our hard work was recognized and we secured first place in the Cohere Track!

The project aims to address the problems associated with academic archive searches. Users often encounter subpar results due to biased ranking and outdated keyword search methods, which can bury relevant papers in deeper pages or exclude them from the results entirely. Additionally, users spend considerable time clicking on articles and attempting to understand their relevance.

Our solution leverages Cohere's semantic embeddings to reorganize search results, making them more relevant and accessible. We input the topic and research into our system, generate refined search criteria for the academic archive, and then query the site. The results are reordered using Cohere rerank and displayed to the user. Moreover, we provide a high-level summary of the selected articles to help users understand their content quickly. Additionally, we are using Anthrophic Claude API's to summarize the whole research paper.

This system also has potential applications in other domains such as patent filing, where it can find similar patents, and law, where it can identify relevant cases. It can even help in searching internal knowledge bases for pertinent information. This innovative solution is designed to make information search more efficient and effective.

Over the course of a semester, we designed, created, and pitched a demo app for Google Meet with mentorship from a team in Google's Corporate Engineering division. This project allowed us to apply our technical skills while incorporating real-world feedback from industry professionals.

Google MeetSense was created to answer the following question:

How might we ensure that meeting organizers have a clear understanding of how to move forward after virtual medium-sized meetings?

Over the course of five months our team has worked to create a prototype application that demonstrates the vision for a tool that could provide a post-meeting summary, action items, and other metrics to meeting organizers. During this time, we have received feedback from a team in Google's Corporate Engineering division, our BigCo studio instructors, peers, and others. We have conducted numerous interviews to guide our thinking toward this demo application. The result is a combination of a Figma mockup, an API that processes an audio file of a real meeting to produce a real meeting summary and statistics, and a front-end application that displays this information.