Ubora-AI: Auditable Commercial Clove Grading & Capture-Session Integrity
Participatory computer vision datasets (CLOVES-4603, ZSTC-Clove-V1), compositional Explanation Evidence Ratio (EER) auditing, capture-session confounding diagnostics, and the offline Ubora-AI Flutter mobile grading app.
Scientific Context & Approach
Zanzibar clove grading directly governs farmer payouts across up to a 53% price gap between Commercial Grade I and lower grades, yet inspection at ZSTC warehouses remains manual and subjective.
In partnership with ZSTC Saateni inspectors, SAAIL Lab released public commercial benchmarks (CLOVES-4603 and ZSTC-Clove-V1), proved in ETRI Journal (2026) that segmentation-first compositional pipelines achieve 100% Explanation Evidence Ratio on true clove tissue, uncovered dangerous capture-session background shortcuts, and shipped the quantized INT8 Ubora-AI Flutter app.
Core Technical Objectives
- Release standardized, multi-grade commercial clove imagery captured under authentic ZSTC grading workflows.
- Enforce compositional segmentation-first inference so grading decisions rely strictly on clove buds rather than tray illumination shortcuts.
- Quantize inference into an offline Flutter + TFLite INT8 smartphone inspector usable without cloud connectivity.
Verified Deliverables & Outputs
- CLOVES-4603 (CVPR 2026 V4A) and ZSTC-Clove-V1 public vision datasets
- Explanation Evidence Ratio (EER = 1.00) compositional classifier published in ETRI Journal
- Offline Ubora-AI Android/Flutter inspection prototype co-designed with Saateni graders