Wireless Capsule Endoscopy (WCE) enables non-invasive imaging of the gastrointestinal (GI) tract. However, it generates over 50,000 frames per exam, making full-frame transmission using radio frequency (RF) communication highly energy-consuming and bandwidth-inefficient. To overcome this, we introduce PeekNet a compact, low-power convolutional neural network tailored for on-device GI lesion classification in a resource-constrained environment. By enabling real-time analysis onboard the capsule, PeekNet allows only clinically relevant frames with high-confidence predictions to be transmitted, conserving energy and extending device longevity. The architecture combines grouped convolutions, parallel dilated branches, and lightweight attention to capture multi-scale context with minimal computational cost. In its fully quantized INT8 form, PeekNet fits within 337kB of flash, executes in under 70ms, and consumes significantly less energy per frame than Bluetooth Low-energy (BLE) transmission. We have evaluated on the Kvasir-Capsule dataset, it achieves 97.24% accuracy using only 0.15M parameters, outperforming conventional lightweight models. This work represents a critical step toward fully autonomous, privacy-preserving, and energy-efficient explainable AI integration in next-generation WCE systems.