EyeRace Pro

The pigeon-racing community has long relied on master breeders to "read the eye" of a pigeon to judge its quality — but that expertise is hard to pass on, judgments drift between experts, and there's no easy entry point for newcomers. EyeRace Pro turns this tacit knowledge into a quantified AI scoring system, delivering an expert-grade evaluation in 30 seconds.

Problem

• Master-level evaluation skills are hard to transfer; the next generation has no efficient path to learn
• Different experts give widely varying scores for the same pigeon — no shared baseline
• 1:1 manual evaluation doesn't scale for breeding farms with hundreds of birds

Solution

A four-step pipeline running entirely inside a Flutter app: Capture → Quality check → ONNX inference → 5-dimension scoring.

1. Smart capture engine: Real-time accelerometer-based shake detection; below-threshold blur or exposure blocks the shot. A three-mode flash auto-adjusts to lighting.
2. Quality check: Post-shot brightness/contrast/blur thresholds; failures auto-retry. Garbage in, garbage out — solved at the source.
3. ONNX inference: A PyTorch-trained EfficientNet-B0 exported to ONNX runs on-device via ONNX Runtime mobile. Average inference < 400ms.
4. 5-dimension scoring: Each eye scored across pupil, inner ring, correlation ring, iris, and outer ring; final weighted recommendation per pigeon.

Stack

• Mobile: Flutter 3.10+, Dart 3, Riverpod 3, GoRouter, Camera + VisionKit, accelerometer
• Backend: Go 1.26 + Gin, GORM, PostgreSQL, JWT, Sign in with Apple / Google
• ML pipeline: PyTorch + EfficientNet-B0, Label Studio annotation, ONNX Runtime mobile
• Billing: StoreKit 2 subscriptions

Outcomes

• iOS / Android MVP shipped, stable for 6+ months
• Backend at eyebird.ailoop.uk, < 200ms average response
• Single-photo evaluation in 30 seconds (capture + upload + inference + render)
• 5-dimension score correlates with senior breeder ratings at r > 0.85

Lessons

• Mobile ONNX has trade-offs: int8 quantization shrinks the model from 50MB to 12MB, but accuracy drops 4%. We landed on a hybrid: cloud inference primary, on-device fallback.
• Capture quality > model quality: The smart capture engine improved overall accuracy more than backend model tuning. GIGO matters at every layer.
• Vertical SaaS wins on depth: Generic OCR or image-classification APIs can't solve domain problems. Going deep is the moat.