This paper presented a tiny ANN-based non intrusive load monitoring (NILM) system designed and deployed on a sub-dollar edge device with extreme resource limitation. We implemented a compact artificial neural network (ANN) model with fixed-point inference entirely on an ESP8266—an ultra-low-cost microcontroller with only 80 KB of SRAM, no operating system, and no hardware floating-point support. The proposed system was trained and optimized using Google Colab, then deployed on the MCU ESP8266 and evaluated through two stages: hardware-emulation experiments, and real-world deployment with Vietnamese household electrical appliances. Using six representative electrical features (Irms, Urms, pf, P, Q, S), our model achieved over 90% multilabel classification accuracy across 127 appliance combinations, with inference latency below 10 ms and a memory footprint under 71.8/80 KB SRAM. Compared to prior works relying on cloud platforms or MCUs with hardware acceleration, our approach enabled scalable, cost-effective, and privacy-preserving NILM deployment at the edge with low resource. Results demonstrated that the Tiny-ANN NILM inference is not only feasible on sub-dollar devices, but also robust to noise and resource constraints, setting a new benchmark for extreme-edge energy disaggregation in smart home and IoT environments.