Card Czar

There is a market for obscure, low-value Magic: the Gathering cards — commons and uncommons from the 1990s that sell for cents — that no store serves. The traditional acquisition process requires conditioning, sorting by expansion, sorting by card number, and integrating with existing stock. For a fifteen-cent card, that labour costs more than the card earns. Card Czar makes the market viable by taking the cognitive work on entirely. The human role is physical: scan cards in, fetch cards when sold. I started buying bulk collections to supply it.

Classification uses a two-stage CNN pipeline in Python with TensorFlow and Keras. A set classifier identifies the expansion, then a per-set network identifies the card number within it. Partitioning the problem this way keeps each network small — sets have at most a few hundred cards — and lets new expansions be added without retraining existing models. Storage is sequential: batch number and position are the retrieval coordinates, so lookup is constant time regardless of inventory size. Pricing runs autonomously against live market data, tracking and undercutting the competitive range without manual intervention.

Only the Card Czar knows where the cards are and only the Card Czar knows what the cards are. There are currently over 72,000 cards listed for sale. I have never heard of most of them.

Visit the shop →