Card Czar

I build Magic: the Gathering cubes from very old sets — collections of roughly 360 cards chosen for competitive balance and historical interest. The cards I want are cheap and common: bulk commons and uncommons from the 1990s that stores have in quantity but can't economically justify sorting through to sell for fifteen dollars. The labour cost exceeded the sale value.

I solved this by making the labour cost mine and making it negligible. I bought collections in bulk, built an AI system to absorb the cognitive work of managing the inventory, and opened an online store that gave other players access to the cards I had been looking for.

The system uses convolutional neural networks, trained with Keras and TensorFlow, to classify scanned card images. Each card is assigned a storage location at classification time. When a card sells, the system retrieves the location. Retrieval is O(1) regardless of inventory size. A data-driven pricing policy, updated continuously from market feeds, maintains competitive prices without manual intervention.

The human role in this system is: scan cards, fetch cards when sold, and enjoy the vintage Magic collection. The AI role is: identify cards, assign locations, price inventory, and track everything. This decomposition is not an accident — it follows directly from the task-decomposition principle studied in the TRAIT research. Each participant does what they are suited for. The human brings physical presence and the judgement needed for edge cases. The AI brings memory, speed, and consistent application of pricing logic across nine thousand cards.

I am frequently surprised by what I have sold, because only the system knows what is in the inventory. This is by design.

Visit the shop →