Projects

A real-money portfolio I have managed myself since I was 10, using a core-and-satellite strategy: high-conviction mega-cap tech held on a ten-year horizon, alongside smaller positions I trim, add to, and rotate. Every buy, sell, and dividend on the dashboard includes my reasoning, including the trades that went badly.

An AI that taught itself the card game Big Two through more than 30 million deals of self-play, built on a DouZero-style deep neural network trained with no human games to imitate. Both the trained network and a second, odds-based engine are available as fully playable opponents.

An app that reads a 14-tile Mahjong hand, typed or photographed, and ranks the closest winning hands in under a second. Each hand is evaluated against more than 21 million possible winning combinations, accounting for jokers, exposed tiles, and every tile already visible on the table.

A model that predicts Bitcoin's four-year price cycle without ever training on Bitcoin, learning the cycle's shape from other cryptocurrencies instead. Asked to reconstruct Bitcoin's full 2014–2025 price history blind, it matched the real curve with an R² of 0.946.

An AI that reads a book series and writes the entire wiki for it: every character, place, and event, with each fact cited to the page it came from. Articles are spoiler-free, hiding anything beyond the reader's progress in the series. One run turned 2,833 pages into 860 cross-linked articles.

A fully playable pinball game built in Scratch, with hand-written physics covering gravity, bounces, and flipper impulses, and a hand-modeled hitbox for every bumper, wall, and obstacle on the board. The game was built in a week and a half for a course on classic video games.