A long-context handling pipeline combining soft activation-space compression, hard token pruning, and KV-cache quantization — relatively nascent research directions — into a unified system for faster long-context inference. Implemented token pruning by fusing two importance signals: LLMlingua-2 (token-level classification) and AttentionRAG (sentence-level attention scoring). Applied Latent Context Language Models (LCLM) with TurboQuant KV-cache quantization for A100 GPUs. Built with PyTorch, Transformers, Modal, and FastAPI.

A personalized agentic AI study assistant that leverages official course information from Canvas LMS and user-uploaded documents. Implemented a multi-agent RAG pipeline using a LangGraph-based AI engine to minimize hallucinations. Designed two Fetch.ai agents: one for OCR of handwritten notes and miscellaneous study material, and another for sending emails to professors via the Gmail API. Built with React.js, Fetch.ai, Iris DBMS, LangGraph, and Flask.

A novel encoder-decoder transformer architecture to convert images of scientific math equations to LaTeX. Developed a DenseNet encoder with 2D positional encoding, paired with a custom GPT-2-based transformer decoder and an extended RoBERTa-based tokenizer. Integrated Weights & Biases to track test metrics including perplexity, BLEU, and F1 throughout training. Built with PyTorch and Transformers.

An agentic AI platform for healthcare providers that streamlines EHR record loading, retrieval, and assessment. Supports ingestion of medical images, EHRs, ClinicalTrials.gov data, and any JSON 1.1 format (including wearables). Features Pulsi, a chat UI powered by a sophisticated RAG pipeline that converts natural language to SQL for patient record queries, interfaces with the ClinicalTrials.gov API for up-to-date trial discovery, and performs similarity search over a vector database of medical images. Built with React, Python, Flask, SQLite, and OpenAI.

Uses AI/ML to enhance disaster response. The Search & Rescue module processes real-time drone footage with YOLO to detect and mark survivors. The Debris Quantifier analyzes post-disaster video to identify material composition and generates heat maps on Google Maps for efficient recycling and rebuilding. Won Best Use of AI/ML Innovation for the Francis Scott Key Bridge Recovery Efforts at Bitcamp 2024. Built with Python, Flask, React, YOLO, Google Gemini API, and MongoDB.

Led the development of PocketPeds through UMD's App Dev Club — an application to improve pediatric healthcare by helping caretakers calculate safe dosages for pediatric OTC medications. Developed a robust backend integrating MongoDB, PostgreSQL, JWT authentication, and the FDA NDC API. Created an intuitive frontend for scanning barcodes, scheduling medications, and viewing dosage history. Built with React Native, Tailwind, Flask, and PostgreSQL.

A gesture and action recognition system that translates hand movements and voice commands into controls for software, operating systems, music apps, and hardware like robotic claws. Uses MediaPipe for real-time hand tracking, maps finger states to desktop commands, and integrates Whisper for voice-to-text. Won the Spark of Genius award at Hack@CEWIT 2024. Built with Python, OpenCV, MediaPipe, PyTorch, Whisper, and Arduino.

An AI-powered video analysis tool that automatically detects events like gunshots, car crashes, and screams — saving law enforcement from manually combing through extensive footage. Users upload video files and receive timestamped logs of detected events with search functionality. The CNN-based audio recognition model was trained with data augmentation and hyperparameter tuning. Won Best Digital Forensics Related Hack at Hoya Hacks 2024. Built with Flask, PyTorch, MongoDB, and Node.js.

A medical records management system that compiles documents across languages into a single database. Automatically extracts information from EHRs, handwritten notes, and prescriptions via image uploads using Google Cloud Vision API, then translates foreign-language documents to English with Google Gemini Pro. Features a natural language-to-SQL query converter so doctors can filter patient info without writing code, plus medical history summarization, notification alerts, and disease prediction from historical records. Built with React, Tailwind, Flask, PostgreSQL, SQLAlchemy, Google Cloud Vision, and Google Gemini Pro.

A comprehensive study tool that extracts LaTeX notes from class recordings and enables students to query all their study material. Implemented Gemini Pro Vision for OCR and conversion to clean LaTeX code, returned as PDF. Employed RAG using LangChain to fetch relevant data from a Pinecone vector database. Built with React.js, Flask, Pinecone, and LangChain. Check out the YouTube video.

An interactive pixel-art game that lets players reshape California's economy using the doughnut economics framework. Players explore 21 sectors spanning social foundations and ecological ceilings, selecting personas — Explorer, Policymaker, Farmer, or Citizen — each unlocking different playable sectors. A custom Canvas 2D engine renders the doughnut visualization, gravity, collision detection, and camera systems, all backed by real California datasets. Built with React, TypeScript, Vite, Canvas 2D, and D3.js. Submitted to TreeHacks 2026.

An AI-powered presentation generator for visual learners. Given a codebase or documentation, Pitcher automatically creates PowerPoint slides with relevant images and data visualizations. Features an interactive presentation mode with AI voiceover — users can interrupt to ask questions, and the tool dynamically generates new slides with source-cited answers. Uses a custom chunking system, IRIS vectorized database for RAG, and MARP/Mermaid for slide rendering. Built with Next.js, TypeScript, Python, and OpenAI. Submitted to TreeHacks 2025.

An AI-powered mental health support platform featuring Apollo, a voice assistant for medication reminders, mood tracking, and real-time support. Includes prescription OCR scanning, sentiment analysis via RoBERTa, and emergency hotline connections. A RAG-powered peer matching system ("Artemis") connects users facing similar challenges. Built with React, React Native, PyTorch, HuggingFace Transformers, Pinecone, MongoDB, and LangChain. Watch the demo.

An AI-powered web app that generates professional, polished resumes in minutes. Users input their details through an intuitive interface, and the system produces well-structured resume content formatted with LaTeX for clean typesetting. Built with React, Flask, GPT-3.5, and LaTeX. Submitted to HackQuest 2024.