Edge Vision, Low Carbon

A retailer in Thika mounted palm-sized edge cameras over fast-moving shelves. The model runs locally (quantized, no GPU) and flags gaps every five minutes. Result: fewer stock-outs, fewer emergency boda-boda runs from the depot, and no constant video uploads to the cloud. Power draw is low enough to ride on a small UPS during outages, so the alerts keep flowing even when the lights blink. Pick hardware with PoE and 5–7W draw, distill your models, and send only events or JSON to the cloud instead of video. Train centrally but infer at the edge to keep latency low and bandwidth cheap. If you want a second win, point one camera at the cashier line to measure queue length; a tiny model can trigger opening another till without streaming any faces to the cloud. This is how you get AI ops without blowing up your energy or privacy budgets.