A sustainable process for valorization of onion waste would need to entail preliminary sorting out of exhausted or suboptimal material as part of decision-making. In the present study, an approach for monitoring red onion skin (OS) phenolic composition was investigated through Visible Near-Short-Wave infrared (VNIR-SWIR) (350–2500 nm) and Fourier-Transform-Mid-Infrared (FT-MIR) (4000–600 cm−1) spectral analyses and Machine-Learning (ML) methods. Our stepwise approach consisted of: (i) chemical analyses to obtain reference values for Total Phenolic Content (TPC) and Total Monomeric Anthocyanin Content (TAC); (ii) spectroscopic analysis and creation of OS spectral libraries; (iii) generation of calibration and validation datasets; (iv) spectral exploratory analysis and regression modeling via several ML algorithms; and (v) model performance evaluation. Among all, the k-nearest neighbors model from 1st derivative VNIR-SWIR spectra at 350–2500 nm resulted promising for the prediction of TAC (R2 = 0.82, RMSE = 0.52 and RPIQ = 3.56). The 2nd derivative FT-MIR spectral fingerprint among 600–900 and 1500–1600 cm−1 proved more informative about the inherent phenolic composition of OS. Overall, the diagnostic value and predictive accuracy of our spectral data support the perspective of employing non-destructive spectroscopic tools in real-time quality control of onion waste.