Anemia remains a major global health challenge, particularly in under-resourced settings where access to laboratory testing is limited. Non-invasive detection methods using computer vision provide a scalable alternative, but many existing approaches either rely on full-face analysis, require manual preprocessing, or depend on computationally heavy architectures unsuitable for real-world deployment. In this work, we propose ConjAnNet, a two-stage deep learning framework that combines U-Net-based segmentation of the conjunctiva with MobileViT for anemia classification. This end-to-end pipeline ensures clinically guided feature extraction and enables deployment on mobile and edge devices. Experimental evaluation on a balanced dataset of conjunctival images from Kaggle demonstrates strong performance, achieving a Dice coefficient of 0.8612 and IoU of 0.7815 for segmentation, and 96.5 % accuracy with an ROC-AUC of 0.987 for classification outperforming baseline CNN and MobileNet models. While these results highlight the potential of conjunctiva-specific analysis for automated anemia screening, the dataset’s lack of demographic diversity and acquisition details represents a limitation. Future work will focus on validation with clinically acquired, diverse datasets and extension toward multiclass anemia grading.