💡 Real-time photo quality coaching with zero-tap audio guidance during inspections

[github-actions[bot]]

Summary

During photo capture in zero-tap inspection mode, on-device image analysis provides voice feedback to ensure photos are usable for Vision AI analysis: "Move a bit closer to fill the frame with brood cells," "Hold steady — capturing now," "Good shot, brood pattern clearly visible." This prevents the frustrating experience of discovering at home during Evening Review that inspection photos are too blurry, dark, or poorly framed for AI analysis — wasting both the field time spent capturing and the AI processing time.

Market Signal

BroodMinder's BeeTV varroa counting system (shipping ~100 units mid-2026) requires precise camera positioning and has validated that guided capture dramatically improves detection accuracy. No beekeeping app currently provides capture-time photo quality feedback. The broader field inspection industry is adopting AI-guided capture as a 2026 trend: Inspectly360 and InspectorData both highlight real-time quality validation. Vision AI analysis accuracy in the PRD targets a 70% interpretable-result rate (FR30) — photo quality is the single highest-leverage bottleneck for hitting that target.

User Signal

Existing idea #497 (real-time brood frame scoring) and #384 (voice-photo contextual annotation) address what to do WITH photos but not how to ensure photos are GOOD ENOUGH to analyze. This is the prerequisite step — without usable photos, neither brood scoring nor annotation can function. The PRD's Vision AI scope note acknowledges the accuracy target is conditional: "If accuracy targets are not met in beta, Vision AI degrades to media-capture-only with manual tagging." Improving input quality is the highest-leverage way to avoid that degradation.

Technical Opportunity

On-device image quality heuristics are lightweight and well-understood — blur detection via Laplacian variance, exposure analysis via histogram statistics, and framing assessment via edge density. These run in < 200ms on modern smartphones without cloud inference. Gemini Nano or Core ML can provide more sophisticated semantic framing analysis ("Is this a brood frame? Are cells visible?") if available on-device. Audio feedback integrates naturally with the existing TTS pipeline used for hive transition announcements in zero-tap mode. The camera capture flow is already planned in Epic 8, Story 8.4 (photo capture and attachment) — adding quality checks is an incremental enhancement to that existing story.

Assessment

Dimension	Score	Rationale
Feasibility	med	Requires on-device image analysis (standard CV heuristics) + TTS integration; no cloud dependency
Impact	high	Force multiplier for Vision AI — improves all downstream analysis accuracy; unique in market
Urgency	med	Should ship alongside or soon after the Vision AI analysis feature (Epic 8) to maximize its accuracy

Adversarial Review

Strongest objection: Adding audio feedback during photo capture could be disruptive in the field — bees are sensitive to vibration and noise. Beekeepers working through multiple hives may find constant feedback annoying when trying to move quickly.

Rebuttal: Audio feedback is brief (2-3 words), played through the phone speaker that's already announcing hive transitions in zero-tap mode. Volume and sensitivity are configurable — advanced beekeepers can disable coaching. The alternative is worse: discovering at home that 5 of 10 photos are unusable costs far more time than a 2-second "move closer" prompt in the field. BroodMinder's BeeTV experience with ~100 production units validates that guided capture is worth the interaction cost when analysis quality depends on it.

Suggested Next Step

Prototype on-device image quality scoring using three heuristics: (1) Laplacian variance for blur detection (threshold: > 100 for acceptable sharpness), (2) histogram spread for exposure analysis (reject images where > 40% of pixels are clipped), (3) edge density for framing assessment (ensure sufficient detail in the central region). Measure latency on target devices (iPhone 12+, Pixel 6+) to confirm < 500ms total analysis time. Define the audio prompt vocabulary (5-7 short phrases mapped to quality issues).