Projects

A deeper look at my recent work.

Minecraft RL Combat Agent

Self-play RL agent that beats 99% of human players in Minecraft sword duels

A reinforcement learning agent that plays Minecraft sword duels, trained with PPO self-play in a headless C++ reimplementation of the game's combat physics. The bot is currently the strongest performing autonomous Minecraft PVP agent in the world. A Java client mod runs the full network inside the real game, where the agent wins about 99% of live duels against human players and stays clean on anticheat flags.

C++
PyTorch
Java
Reinforcement Learning

Demo

Pufferfish

Chess engine that placed 3rd overall out of 75 teams at ChessHacks

Built the Pufferfish chess engine for ChessHacks, placing 3rd out of 75 teams by combining Python-based search logic with an NNUE evaluation function.

Python
Pytorch
Huggingface

Source

LaunchPilot

Multi-agent launch copilot that won 2 tracks at Hack Canada 2026

Built in 36 hours at Hack Canada 2026, winning the Backboard.io Best Use track and placing Top 3 in the SPUR Build a Real Canadian Startup track. LaunchPilot turns a product description into competitor analysis, scored positioning wedges, and a 7-day launch plan by chaining 3 agents through strict JSON schemas. Each stage persists its own agent memory, letting users resume a launch mid-flow, with a Next.js workflow guiding builders from project brief through market research, positioning, launch assets, and approval-gated outreach.

Next.js
TypeScript
FastAPI
Postgres

Source

GooseMarket

Waterloo-only prediction market for campus events

GooseMarket is a Waterloo-only prediction market platform built with Flask and Supabase that lets users trade on campus events using an internal currency. The system includes a complete trade, accounting, and portfolio pipeline optimized for backend efficiency and reliability. An automated market maker is used to provide liquidity for markets.

Flask
Supabase
Python
SQL

Source

SpeedShield

Machine learning model to find optimal speed camera placement locations in Toronto

Parsed and engineered key features from 6 public City of Toronto traffic datasets to create a comprehensive training dataset for our model. We achieved 95% accuracy on validation dataset of existing cameras, demonstrating strong model alignment with existing placement criteria. To present our findings, we created informative visualizations and an interactive frontend to display optimization results.

Python
Scikit-Learn
Matplotlib

Source