Projects

Agent Mahjong

This project evaluates Deep Reinforcement Learning (DRL) agent architectures, benchmarking multi-channel CNNs against Graph Neural Networks (GNNs), to solve the 144-tile Turtle Mahjong layout. To eliminate training noise, Answer Set Programming (ASP) filters out the 30–50% of boards that are mathematically unsolvable, before training begins. To address the temporal credit assignment problem, I am exploring the integration of a C++ binary matrix logic programming module. Depending on performance constraints, this will either run inline within the training loop or as a post-episode processing step to retroactively rewrite the replay buffer—potentially leveraging a binary search algorithm to pinpoint the exact step the agent made the board unsolvable. I audit model behavior, board difficulty, and training progression using a combination of low-fidelity Matplotlib plots and custom Blender scripts to render high-fidelity 3D visual inference replays, in addition to regular training metrics. Full code and data analysis will be posted on GitHub. (Last updated 5/17/2026)

Thesis - “Model-Agents of Change”

This project investigates an interdisciplinary framework for AI innovation, drawing from neuro-symbolic AI, psychology, and systems thinking. I developed a hybrid semantic search system that maps dual-process theory to technical architecture: using semantic embeddings for intuitive “System 1” retrieval and Answer Set Programming (ASP) for structured “System 2” logic. This prototype serves as a technical validation of interdisciplinary conceptual integration, demonstrating how a multi-agent approach can not only leverage the wisdom of other fields, but prioritize modularity and human-centric design over traditional metrics and raw optimization.

WCIW/D Website

During the global pandemic, I led a digital transformation for an international website by migrating its core infrastructure from a restrictive drag-and-drop page builder to a custom-built PHP WordPress plugin. This shift replaced manual workflows with an automated system for collecting and posting global event data, increasing efficiency by over 10x. The new architecture integrated a custom registration engine, a streamlined approval system, and data-driven dashboards for real-time registration insights. To ensure the platform's success on a global scale, I collaborated with an advisory board of creative professionals across six continents, delivering a scalable, high-performance solution that met the urgent demands of a rapidly changing landscape.