Portfolio

Developed in collaboration with Professor Haiyi Zhu, Anna Fang, and Alekhya Maram at Carnegie Mellon University. Simulus is an AR/VR stress-relief simulation platform on the Meta Quest 3, enabling users to practice stress management and navigate everyday high-pressure situations by interacting with smart, AI-driven virtual avatars.

• LLM-Based Virtual Avatars: Built expressive GPT-4o-driven avatar profiles inside Unity, configuring detailed personas, conversational styles, and specialized psychological guidance templates.
• Structured Output Dialogue State: Engineered structured JSON prompting rules to enforce consistent outputs from GPT-4o, facilitating elegant chat parsing and real-time state tracking.
• Speech-to-Text Conversational Core: Integrated OpenAI's Whisper API in C# to achieve seamless, low-latency vocal interactions, letting users communicate naturally using their real voice.

An end-to-end deep learning sound awareness platform designed for Deaf and Hard of Hearing (DHH) individuals. The system continuously listens to environmental sounds and alerts users to critical events—such as fire alarms, sirens, microwave beeps, and doorbells—through localized watch haptics and visual notifications.

• On-Device Sound Classification: Deployed custom audio classification neural networks locally on iOS using CoreML and Apple's SoundAnalysis framework, achieving high-accuracy detection without sacrificing user privacy.
• watchOS Synchronization: Developed a lightweight, highly responsive watchOS companion app using Swift and WCSession to trigger custom haptic patterns matching specific hazard levels.
• Analytics Backend & Dashboard: Programmed a telemetry system with FastAPI, PostgreSQL, and React to help researchers study long-term auditory assistant usage trends securely.

EverArc is a minimalist, highly crafted habit tracking app built for iOS. Born out of a personal frustration with bloated logging tools, EverArc focuses on minimizing friction and delivering stunning, visual rewards that help users stick to their routines.

• One-Tap Seamless Logging: Combines highly responsive, physical-feeling haptic interactions with interactive completion rings to make habit tracking rewarding.
• Native Local Data Architecture: Implemented SwiftData to orchestrate schema modeling and lightweight database operations on-device, offering instant offline loading.
• Streak & Visual Analytics: Created visual calendars and charts with SwiftCharts to track consistency, along with celebratory completion particle effects.

An augmented reality mobile experience designed to help users manage anxious thoughts. Inspired by *Cognitive Defusion* (a core technique in Acceptance and Commitment Therapy), the app enables users to externalize negative thoughts by projecting them into physical space and peacefully letting them go.

• AR Spatial Defusion: Places user-submitted negative thoughts onto floating, customizable 3D text clouds mapped in the room using ARKit.
• Breath-Powered Cloud Dispersal: Implemented live audio classification models via CoreML to detect long exhaling sounds into the phone mic, allowing the user's physical breath to push the AR clouds away.
• Mindfulness Logging: Uses SwiftData to track categories of thoughts and encourage positive reframing patterns over time.

An AI-powered technical document assistant engineered specifically to parse, index, and query the open-source OpenTitan secure silicon platform documentation. It enables silicon designers and software developers to retrieve precise technical specifications through natural language questions.

• High-Fidelity Document Retrieval: Implemented dense document vector databases using FAISS and Hugging Face's sentence transformers, optimizing passage-level semantic search accuracy.
• Context-Aware Synthesis: Built multi-turn conversational agents with Claude 3 Opus and LangChain, enabling detailed hardware reasoning and direct links to official documentation sources.
• Full-Stack Chat App: Wrapped the retrieval pipeline in a lightweight Flask API, serving a gorgeous, responsive, syntax-highlighted React chat dashboard.