💡 Multi-signal swarm prediction engine with seasonal risk calendar

[github-actions[bot]]

Summary

Combine structured inspection observations (queen cells, congestion, drone build-up, bee temperament), seasonal timing, weather patterns, and optional telemetry data to predict swarm probability 7-14 days out. Display as a per-hive risk score integrated into the weekly planning queue and a seasonal risk calendar showing probability ranges by week. Start with rule-based heuristics, evolve toward ML as longitudinal data accumulates.

Market Signal

No competitor predicts swarms proactively — BroodMinder detects swarms AFTER they happen via weight/temperature spikes. 55.6% colony losses in 2024-2025 (worst ever recorded) create urgent demand for preventive tools. Swarm prevention is the #2 concern for hobbyist beekeepers after varroa management. The $12.6B global beekeeping market's growth depends on reducing colony losses, and swarm loss is a significant but preventable contributor.

User Signal

PRD FR17 specifies risk-themed seasonal signals including swarm risk. The weekly planning queue (FR13-FR14) needs risk signals to prioritize inspections. Beekeepers consistently report swarm prevention as a top anxiety driver, especially in spring/early summer. Phase 3 of the PRD envisions "mature predictive intelligence (for example swarm-risk forecasting at higher confidence)" — this proposal defines the concrete path to get there.

Technical Opportunity

Broodly's structured voice-first inspection logging creates a unique data asset: every inspection captures queen cell observations, congestion levels, temperament changes, and drone activity — the classic swarm predictors that beekeepers have used for centuries but never tracked computationally. The Go service layer can compute risk scores from structured observation data. Start with rule-based heuristics (seasonal timing window + queen cell presence + congestion score + days since last inspection + regional swarm timing data), then train ML models as the dataset grows. Optional telemetry (weight, temperature) enriches the signal when available via the sensor-agnostic ingestion layer.

Assessment

Dimension	Score	Rationale
Feasibility	med	Rule-based v1 is straightforward; ML improvement requires longitudinal data that doesn't exist yet
Impact	high	Swarm prevention saves colonies and production; no competitor offers proactive prediction
Urgency	high	Peak swarm season is spring/summer; colony loss crisis makes prevention tools urgent

Adversarial Review

Strongest objection: Without hardware sensors (weight, temperature), the prediction signal from inspection observations alone is too weak and infrequent (weekly at best) to be reliable for 7-14 day forecasting.

Rebuttal: Beekeepers have predicted swarms from manual inspection observations for centuries — queen cells, congestion, and bee temperament ARE the gold-standard swarm predictors. Broodly's advantage is structuring and tracking these signals computationally over time, enabling pattern detection that mental math cannot achieve. Sensor data improves the signal but isn't required for a useful v1. Even an imperfect "elevated swarm risk" flag saves colonies that would otherwise be lost to undetected swarming. The confidence scoring framework ensures predictions are honestly communicated.

Suggested Next Step

Define the rule-based swarm risk heuristic with explicit inputs (queen cell presence, congestion score, seasonal window, days since last inspection, regional swarm timing data from Bee Informed Partnership), score thresholds, and confidence levels. Validate the heuristic against published swarm-prediction research and documented swarm case studies.