Rainfall-related risks can disrupt agricultural activities, while conventional threshold-based warning systems may produce abrupt decisions near predefined boundaries and inadequately represent uncertainty among weather variables. In this study, we developed an Internet of Things (IoT)-based rainfall-risk alert system using a Mamdani fuzzy inference system implemented on a low-cost ESP32 platform. Hourly rainfall, relative humidity, and hourly atmospheric pressure change were used to classify conditions into Safe, Warning, and Danger levels. A publicly available weather dataset recorded at 10 min intervals was aggregated into 8780 complete hourly samples. Because only 67 samples belonged to the Danger class, a conditional Wasserstein generative adversarial network with gradient penalty generated additional 6220 samples for rare-class and boundary-condition stress testing, producing a balanced dataset of 15000 samples. The ESP32 implementation reproduced the MATLAB reference without implementation-induced label differences across the evaluated datasets, indicating the consistent transfer of the Mamdani fuzzy model to a resource-constrained embedded platform. For the combined dataset, the defuzzified outputs produced a mean absolute error of 7.2458 × 10−5 and a root mean square error of 1.3630 × 10−4. A conventional threshold-based baseline achieved 69.987% agreement and a Macro-F1 score of 69.801%. The active buzzer, Message Queuing Telemetry Transport (MQTT), and Telegram channels achieved 100% success during controlled experiments. These findings confirm the implementation consistency and communication feasibility of the proposed prototype, although field-level predictive validity requires further evaluation.