flagship · years in the making
tcgslayer: an AI that plays Pokémon TCG
My deck builder and battle engine, started long before anyone offered prize money for it. Current milestone: the Kaggle AI challenge. Devlogs have the receipts.
The problem
Deck building in Pokémon TCG is a search problem pretending to be a hobby. Thousands of cards, a meta that shifts every set, and the only way to know if a deck works is to play it — dozens of times, against decks you’d have to build too.
I wanted a tool that could build a deck and pressure-test it: an engine where you battle an AI that actually understands energy curves, prize trades, and when to bench-sit a support ‘mon.
What I’m doing
- A battle engine that simulates full games — attach, evolve, attack, prize checks, the boring rules-lawyer parts included
- An AI opponent trained by self-play, so it learns lines nobody hard-coded
- A deck builder that scores lists by simulated win rate instead of vibes
The current milestone
A Kaggle AI challenge showed up offering prize money for exactly the thing I was already building. The engine got a deadline: simulation finals Aug 16, strategy report Sep 14. The devlogs in the logbook track win rates, bugs, and everything the bot does to embarrass me.
Receipts
Live numbers sit on the scoreboard — bot win rate included, updated as the sims run. When it beats me consistently, that counts as shipping.