This site is still being tuned — check back soon for more.
← Projects
Research & Product · 2026

Pokémon Card Scanner & Pricing Tool

Photograph a Pokémon card, get its identity and live market price back — computer vision pipeline with a locally fine-tuned EfficientNet model.

Photograph a Pokémon card, get its identity and live market price back. Detection uses classical CV (Canny edge detection into a perspective warp, HSV fallback, 4-rotation auto-orient) before a two-tier identification pipeline: a vision-language model (Gemini, or a local Ollama model offline) reads name/set/HP directly, falling back to EasyOCR on a normalized crop when foil glare breaks confidence. A locally fine-tuned EfficientNet embedder catches what OCR misses — glare, off-angle shots, foil distortion — by matching against known card embeddings. Matches are priced via the Pokémon TCG API, with the result page showing the raw OCR read for debuggability. Still training and testing locally, not yet deployed.

Django Python PyTorch EfficientNet EasyOCR Gemini Vision API
GitHub Ask about this →
Local Ollama (qwen2.5vl) reading the card — real inference timing: ~19s, 164 tokens/sec prompt eval
Local Ollama (qwen2.5vl) reading the card — real inference timing: ~19s, 164 tokens/sec prompt eval
The full pipeline running — PyTorch/EfficientNet on MPS, then a live Pokémon TCG API price lookup
The full pipeline running — PyTorch/EfficientNet on MPS, then a live Pokémon TCG API price lookup
Running locally — Django + the local vision pipeline
Running locally — Django + the local vision pipeline
Real scan result — N's Reshiram, identified and priced automatically
Real scan result — N's Reshiram, identified and priced automatically
Raw input — handheld photo, no preprocessing
Raw input — handheld photo, no preprocessing
After detection, orientation correction, and crop
After detection, orientation correction, and crop