← Back

Democratizing High-Quality LLMs through Iterative Refinement Pipelines

I'm currently working on developing an optimal iterative refinement pipeline for open source LLMs with the mentorship of members from the NYU Agentic AI Lab. This project was started in the NYU GSTEM program.

Through this project, I hope to democratize access to high quality LLMs and enable user-data transparency while using LLMs. The work focuses on creating systems that allow for continuous improvement of model outputs through structured feedback loops.

I will be presenting my work at the International Conference for AI Research in late December.

View Summer Program Paper →

Similarity Matching Methodology

Previously worked on creating a methodology for similarity matching inspired by ArCADia, a project I proposed during the Conrad Challenge. This project was done under the mentorship of a professor at Texas A&M.

The research focused on improving geometric retrieval efficiency and precision in 3D model comparison tasks.

The system utilized voxel-based sampling and probabilistic shape descriptors—such as D2 and A3 histograms—to encode model geometries into compact, rotation-invariant feature vectors. These representations were indexed using FAISS to enable high-speed similarity search across large CAD datasets. The resulting methodology achieved consistent performance in identifying structurally and spatially related models, offering applications in design automation and computational spatial analysis. For its effectiveness, this project was awarded Best Automation Project at Science and Engineering Fair of Houston out of ~900 projects.

Currently, I am working on implementing this methodology into a Blender extension for 3D modelers to use.

Machine Learning for Inverse Problems in Image Classification

Investigated how machine learning can solve inverse problems for image classification under the mentorship of Andreas Mang and the SCOPA group. The work explored computational techniques for reconstructing input features from classification outputs.

Findings were documented in a comprehensive final report detailing the mathematical foundations and practical applications of inverse problem solving in computer vision.

View Student Paper →

Data Science & Programming Excellence

For demonstrated aptitude in programming and data-science, I was awarded the $6,000 Winston Data Scholar Award from the Winston Data Foundation.

This recognition acknowledges contributions to data-driven research methodologies and computational problem-solving techniques across various projects.