💡 AI-powered colony loss post-mortem with longitudinal root-cause analysis

[github-actions[bot]]

Summary

When a beekeeper reports a colony loss, Gemini analyzes the full inspection trail — observations, photos, treatments, recommendations, and environmental context — from the preceding 90 days to generate a structured learning summary. The post-mortem identifies what signals were present, what actions were taken or missed, seasonal context at the time, and generates specific recommendations for preventing recurrence. Framed as recovery-oriented learning — "here's what we can learn" not "here's what went wrong" — aligned with Broodly's emotional design principles.

Market Signal

Annual colony loss rates are 30–50% for hobbyist beekeepers. No beekeeping app offers systematic learning from colony losses — beekeepers currently piece together clues manually in forums and Facebook groups. Climate-driven changes are increasing loss risk: the 2026 swarm season started nearly two weeks earlier than the previous year (source), and unpredictable nectar flows from sudden cold snaps create food shortage risks. The growing AI pest detection market ($1.2B in 2026, source) demonstrates demand for AI-assisted biological analysis.

User Signal

Broodly's emotional design specification mandates a "recovery-oriented" tone: "reduce fear before introducing complexity" and "warnings are constructive, recovery-oriented, with clear next steps." The PRD includes a Trust Recovery edge case (Hannah follows a recommendation, colony suffers, she returns angry) where "the app does not hide or excuse the outcome — it explains the decision basis transparently and provides a concrete recovery path." Colony loss is the most emotionally devastating event for beekeepers — exactly where this design philosophy should materialize. Existing Discussion #383 (decision-outcome attribution journal) captures outcomes longitudinally but doesn't specifically address the colony loss learning workflow.

Technical Opportunity

The architecture already captures the complete data trail needed: inspection observations with structured_data JSONB, recommendation traces (FR50, FR51), media with AI analysis results, and audit events. Gemini 3.5 Flash's multimodal capabilities can analyze text + image + audio inspection history in a single prompt. pgvector semantic search can find similar loss patterns across anonymized aggregate data, enabling "colonies in your region with similar histories experienced..." insights. The recommendation contract (action + rationale + confidence + fallback) naturally extends to post-mortem learning summaries.

Assessment

Dimension	Score	Rationale
Feasibility	high	All required data already captured; Gemini multimodal analysis of existing inspection records; minimal new infrastructure
Impact	med	Emotionally resonant and unique in the market, but niche use case (triggered only on colony loss events)
Urgency	low	Post-MVP feature; requires accumulated inspection history to generate meaningful analysis

Adversarial Review

Strongest objection: AI root-cause analysis on complex biological systems could be wildly wrong. Presenting an incorrect post-mortem undermines trust exactly when the user is most emotionally vulnerable — turning Broodly's "recovery-oriented" promise into a liability.

Rebuttal: The feature presents observations and patterns, not diagnoses. "Here are the signals we found in your data" — not "here's what you did wrong." Every observation includes a confidence level, consistent with the recommendation contract. The alternative — learning nothing from a colony loss — is strictly worse. The post-mortem explicitly includes a "what to watch for next time" section that builds competence rather than assigning blame. Expert review of the post-mortem prompt template ensures AI outputs are framed constructively. This feature directly embodies the PRD's Trust Recovery edge case: transparent explanation with a concrete recovery path.

Suggested Next Step

Design the colony loss reporting flow (voice-triggered: "Hive 3 is dead" or manual via hive status update). Define the post-mortem analysis prompt template for Gemini with explicit guardrails against blame framing. Create the structured output schema for learning summaries: timeline of observations, detected patterns, environmental context, confidence-scored contributing factors, and actionable prevention recommendations.