Diabetes, a non-communicable disease that damages insulin utilization, requires patients to regularly monitor at home and maintain their blood glucose values within a specific range. The discomfort and infection risks associated with traditional finger-prick methods emphasize the need for noninvasive alternatives. This study proposes an accurate and portable device based on multi-spectral sensing and AI-driven Non-invasive and self-monitoring blood glucose at home. The study evaluated three machine learning models, with the dimensionality reduction of the dataset using Principal Component Analysis. The in vivo testing results are promising, with the Neural Network regression model, using the eight principal components, achieving high prediction performance with the 0.86 coefficient of determination of and the 7.37 mg/L root mean square error. Moreover, the highest clinical accuracy regression model with Clarke Error Grid Analysis has a 100% score in the most reliable zone A. Our portable system offers promising accuracy comparable to invasive methods, enhancing patient comfort and convenience.