Key Features

• 📷 Multi-card batch scanning
• Photograph a whole spread of cards; each one is detected, identified, and streamed back to the UI one-by-one as it resolves (first card in ~3s on a Samsung S22+).
• 🎥 Live continuous scanning
• Hold cards up to the camera and they're identified automatically into a running session list - no shutter tap, with a two-frame confirmation gate to reject phantom matches mid-swap.
• 🤖 Three-mode recognition engine
• OCR text search, CLIP image matching across 47,442 card vectors, or both fused via Reciprocal Rank Fusion for the most accurate identification.
• 🇯🇵 Automatic Japanese support
• A single photo of mixed English and Japanese cards is handled with no toggle; JP names are translated and matched against 27,255 first-class JP records (all eras, 1996–present).
• 💰 Live pricing & sales
• Ungraded and PSA 7–10 prices, recent eBay/TCGPlayer sales, and trend charts scraped from PriceCharting and cached in Redis, with an instant USD/JPY toggle.
• 🔍 Fuzzy search & disambiguation
• Partial names, set hints, and exclusion terms, with trigram matching to recover OCR misreads and a printed-total lookup to separate cards that share a name and number across sets.

• 🎯 Finding cards in a messy photo
• Problem: my first detector inferred card positions by clustering OCR text - it failed completely on holofoil, face-down, or overlapping cards with no readable text.
  Action: trained a YOLO11n detector on 1,688 real images, then built a synthetic-data pipeline that pastes card art onto random desk and display-case backgrounds with auto-generated bounding boxes (zero manual labeling).
  Outcome: detection became robust to surface finish and overlap, and bounding-box accuracy (mAP50-95) rose from 0.904 to 0.964 - giving the matcher cleaner crops.
• 🧠 Too little real data to train the visual matcher
• Problem: an off-the-shelf image model gave near-random matches (it had never seen a phone photo of a card next to its clean scan), and I didn't have enough real labeled photos to fine-tune it.
  Action: generated synthetic training pairs instead - each clean card image augmented with perspective warp, color jitter, blur, JPEG artifacts, and random backgrounds to mimic real scanning conditions - and fine-tuned CLIP's visual encoder on them.
  Outcome: recognition held up across different lighting and angles, lifting match scores from a near-random 0.45–0.58 to 0.78–0.86 on distinctive cards.
• 🇯🇵 A broken shortcut for Japanese cards
• Problem: I first derived Japanese card images by mapping JP numbers onto English sets - but EN and JP numbering are unrelated, so every JP match returned the wrong artwork.
  Action: re-architected the data model, scraping 27,255 JP cards as independent records with their own images, numbers, and sets, and made every search filter by language.
  Outcome: fully independent EN and JP pipelines - mixed-language scans now resolve each card to the correct artwork automatically.
• ⚡ Three slow round-trips per scan
• Problem: a multi-card scan made three sequential network calls (detect → OCR search → image search), so the UI only updated after all three finished.
  Action: collapsed them into a single /scan endpoint that batches all CLIP embeddings in one pass, runs database and OCR searches concurrently, and streams results as NDJSON.
  Outcome: the first card now appears while the last is still processing, instead of waiting for the whole batch.
• 🎥 Live scan too slow to feel real-time
• Problem: the original live mode used a two-step loop with a stability hold - 2–3 seconds per card, hitting the backend twice each cycle.
  Action: rewrote it as a single round-trip (one snapshot → one backend call handling both detection and recognition), removing the on-device step entirely.
  Outcome: cycle time dropped to ~700–900 ms, making continuous hold-and-scan feel instant.