💡 AI-powered phone-camera Varroa mite counting with closed-loop treatment decision support

[github-actions[bot]]

Summary

Enable beekeepers to photograph sticky boards or brood frames during inspections and receive instant AI-powered mite counts with infestation percentages and confidence scores. Unlike standalone counting tools (HiveLog AI, BeeScanning, VarroDetector), this integrates directly into Broodly's recommendation engine to produce treatment decisions with action + rationale + confidence + fallback — closing the loop from observation to decision in one zero-tap voice-triggered flow.

Market Signal

Phone-camera varroa counting is now production-ready across multiple competitors: HiveLog AI counts mites from brood-frame photos with confidence scores, BeeScanning uses phone camera AI, and VarroDetector offers free deep-learning sticky board analysis. BroodMinder is shipping automated hardware-based counting via BeeTV (~100 units mid-2026). However, no competitor integrates mite counts into a decision-support engine that produces contextualized treatment recommendations considering colony history, regional resistance patterns, and seasonal timing. The counting market is fragmenting into standalone tools; the integration opportunity remains open.

User Signal

Treatment timing is the #1 decision anxiety for beginner beekeepers (PRD user journey: Hannah must decide whether to feed, treat, or wait). Varroa destructor is the single largest driver of colony losses. The PRD's recommendation contract (action + rationale + confidence + fallback) is architecturally ready to consume mite count data as a high-weight input signal. The existing Vision AI pipeline (Vertex AI) in the architecture provides the inference infrastructure.

Technical Opportunity

The architecture already includes Vertex AI for inspection photo analysis (architecture.md: Gemini Vision for inspection photo analysis), media upload via signed URLs to Cloud Storage, and async processing via Pub/Sub. The recommendation engine's ExplainableRecommendation contract can consume mite count data directly. On-device inference via ExecuTorch (already tracked in Discussion #382) could enable offline counting. Voice command "check mites" triggers camera → AI count → recommendation in one continuous zero-tap flow.

Assessment

Dimension	Score	Rationale
Feasibility	high	Proven technology (multiple competitors ship counting already). Vertex AI Gemini Vision handles image analysis. Integration with existing recommendation engine extends current contracts.
Impact	high	Directly addresses the #1 decision anxiety (treatment timing) and #1 colony threat (Varroa). Closes the decision loop no competitor closes.
Urgency	high	Multiple competitors shipping standalone counting. Integration opportunity narrows as counting becomes commoditized. First to close the decision loop wins.

Adversarial Review

Strongest objection: Phone-camera counting accuracy varies with photo quality, lighting, and sticky board condition. If the count is wrong, the treatment recommendation is wrong — potentially worse than no recommendation at all.

Rebuttal: This is exactly why the confidence score matters. The system communicates confidence explicitly ("Based on this photo, I estimate 3-5 mites per 100 bees with MODERATE confidence"). Low-confidence counts produce conservative, safety-first recommendations with fallback guidance for manual verification. The recommendation engine already handles confidence-aware fallbacks — this is a core architectural principle, not an afterthought. Users always see the evidence and can override.

Suggested Next Step

Create a technical spike story evaluating Vertex AI Gemini Vision accuracy on sticky board and brood frame photos. Define the integration contract between the Vision AI mite count output and the recommendation engine's treatment decision input. Benchmark against HiveLog AI's published accuracy metrics.