When on-device AI beats cloud for environmental sensors
Edge inference is justified when one or more hold: no connectivity (forest, sea, remote farms); data too big for the link (LoRa carries tens to ~200 bytes per packet, so send class + confidence, never audio or video); battery (the radio costs far more energy than inference, so classifying locally and transmitting only events extends life from days to months); latency (response needed within seconds); fleet scale (many $10 nodes vs per-stream cloud cost). Projects that hit several at once (e.g. acoustic logging detection, camera traps that discard empty triggers, pest-trap counters, bycatch cameras at sea) make the strongest edge-AI pitches; a single reason is weak. Laptop or phone in airplane mode running the same int8 TFLite model is an honest stand-in when hardware is missing.