My research poster for my internship at IoT4Ag, where I developed a framework to diagnose failures in Vision-Language Models (VLMs) for autonomous precision agriculture missions executed by robots.

• Abstract: We employed a VLM to operate post-failure, producing structured, GPS-tagged diagnostic event reports that explain operational failures experienced by the agricultural robot during deployment. The result is a system that passively builds a detailed record of the robot’s difficulties during tasks paired with recommendations over the course of a mission, bridging the gap between robot perception and farmer decision-making.

An independent NLP research paper where I designed and implemented a RAG framework that performs graph-traversal over hierarchical knowledge graphs to deliver structured context to LLMs.

• Evaluated across three world scales (18-800 nodes) and four query types against a baseline using Mistral 7B
• HG-RAG achieved 1.86 factual accuracy vs. 0.02 baseline on large-scale graphs and demonstrated exceptional scaling on multi-hop reasoning (4.10 vs. 1.66 LLM judge score at large scale)

Autonomous Driving Project in BeamNG Tech Simulator to develop fluency in perception, scene-reconstruction, and autonomous model techniques used by modern autonomous-vehicle companies

• Designed and implemented a multi-modal sensor harness over BeamNG.tech, producing synchronized camera, LiDAR, depth, and pose streams for downstream perception research
• In process of reconstructing high-fidelity 3D scenes from multi-camera driving footage using 3D Gaussian Splatting

FlowFundAI is a SaaS finance web app for young adults and college students my group developed for our Full-Stack Web Dev Final Project. Designed with college students in mind, the application bridges the gap between raw financial data and actionable guidance through natural language interaction with Pluto AI.

• Parses financial documents for individual transactions and automatically pushes them to the app dashboard with Google Gemini API
• Features pluto AI assistant, powered by Claude Sonnet 4.6 refined through prompt engineering, who has the ability to access your transaction history and spending habits for personalized guidance
• Developed with React/Vite and Flask; Our team Implemented JWT-based authentication, SQLite data persistence, and a dynamic transaction dashboard

My team and I built an AI-powered transit planning web app for UC Merced students using Google Gemini 2.5 Flash to process natural language queries and recommend CatTracks bus routes.

• Integrated Google Maps Geocoding, Directions, and Distance Matrix APIs to calculate accurate bus paths and walking directions to bus stops
• Supported 9 bus routes with real-time schedule parsing and time-based trip planning given a user-specified arrival time

Project of the Machine Learning Club at UC Merced. Analyzes a user's resume and provides specific feedback with AI; follows up with a list of jobs relevant to the user. I primarily worked on the API calls in the backend and various UI elements.

• Calls ChatGPT API with detailed prompt and user's Resume parsed into text, receives structured feedback
• Calls JSearch API and sends a JSON file of keywords to receive relevant jobs
• Displayed to a user on a website made with React JSX