💡 Expo-AI-Kit powered offline-complete intelligence stack for field-independent operation

[github-actions[bot]]

Summary

Leverage the newly released expo-ai-kit SDK to build Broodly's entire field intelligence stack for offline operation: on-device voice transcription (Whisper), observation classification (lightweight NLP), brood pattern analysis (on-device vision), and recommendation pre-computation using cached models. This makes Broodly the only beekeeping app that provides AI-powered recommendations, voice logging, AND photo analysis without ANY internet connectivity — critical for the estimated 40%+ of apiary locations with no cell service.

Market Signal

HiveSense offers offline-first operation with on-device voice transcription but no offline AI recommendations or photo analysis. Apis App's voice AI appears to require connectivity for full NLP classification. expo-ai-kit launched in 2026 as a production-ready SDK supporting LLM, STT, vision, and embeddings on-device for Expo apps (source) — eliminating the need for custom ExecuTorch integration. Edge AI research published in 2026 shows agriculture pest detection models achieving 98.6% accuracy at just 1.2MB model size (Tiny-LiteNet, source). React Native ExecuTorch enables Whisper Tiny (150MB total) on-device. This is a technology inflection point that makes full offline AI feasible for mobile field apps.

User Signal

The PRD specifies cache-first reads and local media staging as MVP requirements, with full offline sync deferred post-MVP. However, "no offline writes" means that in poor connectivity areas, the app's AI capabilities are severely degraded — exactly when the beekeeper needs guidance most. Discussion #346 (on-device voice transcription) and Discussion #382 (ExecuTorch full-stack inference) validate strong interest in offline AI but predate expo-ai-kit's availability, which significantly lowers the implementation barrier compared to raw ExecuTorch.

Technical Opportunity

expo-ai-kit provides a unified SDK for on-device inference across speech, vision, text, and embeddings — directly compatible with Broodly's Expo stack. This is significantly simpler than the ExecuTorch approach in Discussion #382 because expo-ai-kit handles model management, memory allocation, and platform abstraction. Key specs: Whisper Tiny at 150MB is viable even on mid-range phones. Brood pattern classification can use lightweight models (1-5MB) with transfer learning from agriculture pest detection architectures. The SpokenAudio mode includes noise suppression optimized for outdoor voice capture — directly applicable to beeyard conditions.

Assessment

Dimension	Score	Rationale
Feasibility	med	expo-ai-kit is production-ready for Expo; main risk is model accuracy vs cloud Gemini and battery management
Impact	high	Only beekeeping app with full offline AI; addresses the fundamental connectivity challenge in field beekeeping
Urgency	high	HiveSense already proves offline-first with on-device AI is a winning proposition; expo-ai-kit makes parity achievable

Adversarial Review

Strongest objection: On-device models are less accurate than cloud models (Gemini). Large model sizes could consume storage on low-end phones. Battery drain during inference could leave the beekeeper with a dead phone mid-inspection. The approach risks delivering degraded AI quality that damages user trust.

Rebuttal: expo-ai-kit handles model optimization and quantization automatically. Whisper Tiny at 150MB is comparable to a few photos in storage — acceptable on any modern phone. Battery monitoring can trigger graceful degradation (switch to recording-only mode if battery drops below 20%). The key insight: degraded AI in the field is infinitely better than NO AI in the field. Cloud models remain the gold standard when connectivity is available — on-device is the offline fallback, not the replacement. Frame it as "AI everywhere, best AI with connectivity" rather than either/or. HiveSense's rapid growth in 2026 proves that offline-first with on-device AI is what field users actually want.

Suggested Next Step

Install expo-ai-kit in a test branch and benchmark three critical capabilities: (1) Whisper Tiny transcription accuracy on beekeeping-specific vocabulary, (2) a fine-tuned MobileNet brood pattern classifier inference time on mid-range Android, (3) total memory and battery impact of running STT + vision concurrently during a simulated 30-minute inspection session.