Accurate quantitative measurement of drugs and their metabolites is important as this can be used to establish long-term abuse of illicit materials as well as establish accurate drug dosing for legal therapeutics. However, the levels of drugs and xenometabolites found in human body fluids necessitate methods that are highly sensitive as well as reproducible with the potential for portability. Raman spectroscopy does offer excellent reproducibility, portability and chemical specificity, but unfortunately, the Raman effect is generally too weak unless it is enhanced. We therefore developed surface-enhanced Raman scattering (SERS) and combined it with the powerful machine learning technique of artificial neural networks to enable rapid quantification of caffeine and its two major metabolites theobromine and paraxanthine. We established a three-way mixture analysis from 10−5 to 10−7 mol/dm3, and excellent predictions were generated for all three analytes in tertiary mixtures. The range we selected reflects the levels found in human body fluids, and the typical errors for our portable SERS analysis were 1.7 × 10−6 mol/dm3 for caffeine, 8.8 × 10−7 mol/dm3 for theobromine and 9.6 × 10−7 mol/dm3 for paraxanthine. We believe this demonstrates the exciting prospect of using SERS for the quantitative analysis of multiple analytes simultaneously without recourse to lengthy and time-consuming chromatography, a method that often has to be combined with mass spectrometry.