Current
An agent harness that transforms user feedback from X/Twitter or SMS/calls into sandbox-tested, preview-ready pull requests. Project for TreeHacks 2026.
An agent harness that transforms user feedback from X/Twitter or SMS/calls into sandbox-tested, preview-ready pull requests. Project for TreeHacks 2026.
A multi-stage RAG architecture for financial document question answering that combines document-level routing, dense retrieval, cross-encoder reranking, and reduced-context LLM querying. Improves accuracy from 72% to 92% on financial QA tasks while reducing token costs by 100x compared to raw LLM processing.
Research investigating dynamic token pruning in autoregressive transformers to reduce computational costs while maintaining generation quality. Combines Learned Token Pruning (LTP) with reinforcement learning-based Dynamic Token Reduction to adaptively remove unimportant tokens during inference.
A video editor that automatically creates highlight montages by syncing audio beat drops with key moments in your footage. Upload your clips (basketball, gaming, etc.) and audio - the algorithm handles the editing to create seamless highlight reels.
A mobile app for students to browse school clubs, join activities, and track their schedules through a personalized calendar. Club leaders can take attendance and make announcements. Built as my introduction to full-stack development with authentication and database features, presented at the Bronx Borough Science Expo.
The official site for the Bronx Science National Honor Society (NHS) provides details on the application process, member responsibilities, and tutoring forms. Created as my first web development project, it is now used and maintained by student leaders at Bronx Science.
A fully animated project created in Autodesk Maya. All 3D models were built and rendered in Maya, with Python scripts used to automate repetitive tasks like randomizing DNA helix and cell positions. Created as a final project for MITES Semester 2023.