This paper presents a lightweight and robust face recognition system designed for deployment on resource-constrained embedded platforms. Although MobileFaceNets provide an efficient baseline for mobile face recognition, our analysis shows that the model still contains redundant channels that limit the consistency performance. To address this issue, we introduce a structured pruning strategy that evaluates channel importance using \(\ell _2\) -norms and cosine similarity, allowing us to remove low-contributing or highly correlated filters while preserving the depthwise separable structure. To enhance robustness under real-world conditions, we incorporate a stereo-camera acquisition setup, in which two synchronized cameras capture complementary facial views. Each image is processed by an identical lightweight model, and the resulting embeddings are fused to improve feature stability and reduce errors caused by pose or illumination variations. The complete system–including detection, alignment, pruned MobileFaceNets backbones, and deployment on a Raspberry Pi 5 with a Hailo-8L AI accelerator. Experiments on a dataset of 9 faces demonstrate that the proposed approach significantly reduces the number of parameters while enhancing reliable recognition accuracy, highlighting the effectiveness of combining model pruning with stereo-based feature fusion for embedded face recognition.