My Projects

Architected and deployed a full-stack Fallout-themed web application with dynamic responsiveness and custom animations. Implemented 10+ interactive terminal features using Svelte/TypeScript frontend and LAMP stack for backend. Built a comprehensive RESTful API for database communication, including account management, forms, notes, and settings. Features a state management system using Svelte stores and cookie synchronization, enabling persistent user customization across 12+ configurable settings with seamless cross-session experience.

Developed a persona-driven conversational AI using adapter-based fine-tuning on Phi-2 Instruct model with PyTorch and HuggingFace Transformers. Implemented PEFT (Parameter-Efficient Fine-Tuning) architecture with modular checkpoint management system, enabling rapid persona deployment across multiple model iterations. Architected scalable foundation for RAG integration with robust prompt engineering framework and output optimization. Includes analytics dashboards through Power BI to monitor user interactions, token usage, sentiment patterns, and failure modes.

Developed a CNN ensemble model for melanoma detection achieving 95% accuracy on the ISIC 2024 challenge dataset. Combined multiple architectures with specialized loss functions and class weighting to reduce false negative rates by 85%. Engineered a robust data pipeline processing 300GB+ of medical imaging data with advanced augmentation techniques including rotations, flipping, and color jittering. Integrated 2024 and historical datasets for comprehensive training coverage, demonstrating strong performance in real-world medical imaging scenarios.

Built an AI-powered thumbnail generator using U-Net architecture and Stable Diffusion, trained on Azure GPU compute environments for scalable experimentation. Engineered an end-to-end deep learning pipeline with modular Python architecture encompassing preprocessing, training, and inference scripts. Implemented Azure Blob Storage-backed dataset pipelines to manage 10,000+ images (50GB+) with automated ingestion, augmentation, and metadata tracking via CSV. Integrated Azure DevOps-style CI/CD practices for model versioning and reproducible experiments.

Built a pure C++ neural network trainer achieving 97-98% accuracy on MNIST with real-time visualization and interactive testing interface. Optimized the training pipeline with loop unrolling and batch processing, reducing training time by 200%. Architected a multi-threaded visualization system with SFML, enabling real-time network monitoring during training. Features include from-scratch implementation of backpropagation, gradient descent, and various activation functions, demonstrating deep understanding of neural network fundamentals.

Designed and deployed a web-based SQL game using HTML, CSS, JavaScript, and PGLite, gathering over 5 million visits. Developed 10 advanced SQL challenges, improving user engagement and platform growth. Enhanced website speed and performance via schema optimizations and additional software changes by 60%. The platform provides an interactive learning experience for SQL practitioners of all skill levels, with real-time query execution and immediate feedback.

Undergraduate thesis research developing a neurosymbolic memory architecture for identity-based AI applications. Implements dynamic knowledge graphs that enable LLM agents to recursively construct and evolve belief structures through multi-turn interactions. Features AGM belief revision theory for graceful contradiction resolution, confidence-based epistemic entrenchment preventing catastrophic forgetting, and metacognitive monitoring for self-aware reasoning. Combines symbolic graph structure with neural semantic embeddings, enabling hybrid retrieval that merges RAG's intuitive recall with structured logical reasoning. Addresses critical gaps in companion AI, game NPCs, and therapeutic agents by providing persistent memory, coherent personality simulation, and temporal belief evolution. The architecture demonstrates how LLMs can achieve genuine continual learning without gradient updates through experience-driven belief formation.

ConUHacks X hackathon project building an end-to-end fitness gamification platform powered by Self-Identity Knowledge Graphs (SIKG). Transforms real-world fitness progress into a personalized fantasy RPG where AI generates unique worlds and adaptive quests calibrated to user goals. Built with React/TypeScript frontend (TailwindCSS, shadcn/ui, motion/react animations) communicating via REST API with FastAPI backend (Python, Claude AI integration, Supabase PostgreSQL). Features include: AI-generated unique fantasy scenarios, 2-10 daily personalized fitness tasks, optional quests with stat requirements, reflection authenticity grading via LLM, knowledge graph tracking behavioral patterns and personality insights, and character stat progression based on real training patterns. Solved complex challenges including SIKG implementation from first principles, multi-model AI integration (Haiku/Sonnet/Opus), knowledge graph edge generation, and full-stack integration across async operations. Built in 24 hours by separating backend narrative engine from frontend UI, with AI-assisted development allowing focus on features and prompt engineering rather than boilerplate. Shipped a working product demonstrating novel gamification approach where self-improvement becomes genuinely fun through meaningful storytelling.