CardFlux
AI-powered trading card identification
The Challenge
Card shop owners spend 3-5 minutes per card manually identifying and pricing inventory. With thousands of cards flowing through shops weekly, this creates a massive bottleneck. Existing solutions were slow, inaccurate, or required expensive specialized hardware.
The Solution
I developed a desktop application using Electron and React that leverages DINOv2 for visual embeddings and ORB for geometric verification. The multi-modal scoring system combines neural network features with traditional computer vision for 100% accuracy. A persistent Python ML bridge enables model reuse across identifications.
Key Features
111ms Identification
Achieved sub-200ms card identification through DINOv2 FP16 inference, pre-computed ORB keypoints, and parallel geometric matching.
100% Accuracy
Multi-modal scoring combines visual embeddings (DINOv2) with geometric verification (ORB features) using dynamic weight allocation.
5,390 Card Database
Full One Piece TCG coverage with 1,014 reprints/variants automatically grouped. Extensible to Pokemon and Magic: The Gathering.
Shop Management SaaS
T3 stack web platform with Clerk auth, multi-org support, inventory tracking, and real-time price sync from TCGPlayer.
The Outcome
Achieved 111ms identification time with 100% accuracy across 5,390+ cards. Cold start optimized from 10.5s to 2.3s (78% faster). The system processes the entire One Piece TCG catalog including 1,014 reprints and variants, with architecture extensible to Pokemon and Magic: The Gathering.
Engineering Highlights
- Built persistent Python ML bridge with JSON-RPC communication for model reuse across identifications
- Optimized cold start from 10.5s to 2.3s (78% faster) through model preloading and warmup inference
- Pre-computed 120MB ORB keypoints cache eliminates 60% of geometric computation time
- Designed monorepo with pnpm workspaces spanning desktop app, web app, and data pipeline services