We describe a novel, portable optical tool enabling non-invasive hematocrit determination from a single drop of blood. An attractive alternative to current methods, this fast, precise, and portable solution is particularly adapted to ex-vivo clinical environments such as dialysis or mobile clinics. The NIR spectrum of the term is measured by the instrument, which is subsequently preprocessed for noise removal through various filters like Standard Normal Variate (SNV), Multiplicative Scatter Correction (MSC), and smoothing through the Savitzky-Golay filter. To develop supervised learning models, the preprocessed spectra are combined with hematocrit values that had been measured. Training and test data sets were split at 70% and 30%, respectively. Hyperparameters of the Long Short-Term Memory (LSTM) and Support Vector Machine (SVM) models were optimized by cross-validation. Satisfactory results were obtained on tests with determination coefficients (R2) 0.914 for SVM and 0.929 for LSTM. The stability of the models was verified in a second test on 80 new observations with even better performance (R2 = 0.972 for LSTM, 0.956 for SVM). Therefore, from the simple blood test, this method can reliably estimate hematocrit values. Advanced AI algorithms and a compact unit offer the possibilities for handheld, accessible, and efficient medical therapies in various clinical applications.