Banana cultivation plays an important role in the agricultural economy of Andhra Pradesh, particularly in the Ravulapalem region, which is well known for banana production. This study presents EAAF-Net (Ensemble AI Annotation Framework), an applied AI workflow for banana lifecycle classification and semi-automated image annotation using images acquired with a Raspberry Pi-based imaging system under controlled postharvest storage conditions. The workflow is designed to support the classification of three banana lifecycle stages—unripe, ripe, and spoiled—while reducing manual annotation effort and facilitating the development of reliable agricultural image datasets. Rather than proposing a new deep learning architecture, EAAF-Net combines three established pretrained convolutional neural network backbones, namely, EfficientNet-B0, Xception, and MobileNetV2, through ensemble probability averaging. A confidence-guided decision mechanism automatically accepts high-confidence predictions while identifying uncertain samples for manual verification, thereby improving annotation consistency and supporting efficient semi-automated dataset construction. The workflow is implemented as an edge-artificial intelligence (AI) system in which image acquisition and model inference are performed locally, without cloud computing or network-based Internet of Things (IoT) connectivity. The evaluation demonstrates that the proposed workflow provides competitive classification performance while offering a practical and scalable framework for confidence-guided annotation of banana lifecycle images. The proposed approach provides a useful foundation for intelligent vision-based applications in postharvest quality assessment, banana lifecycle monitoring, and precision agriculture, with future work directed toward broader field validation and IoT-enabled agricultural monitoring systems.