There are currently on-sky spectrographs with sub-10-cm/s precision. One such instrument is EXPRES, a R~137,000 optical spectrograph that has demonstrated 4-7 cm/s instrument precision. I will discuss ways machine learning can be used on this intensely stable data to acquire more precise spectra and cleaner RVs. I will highlight excalibur, a novel hierarchical, non-parametric method for wavelength calibration. With EXPRES data, using excalibur-generated wavelengths reduced the overall RV RMS of several data sets by 0.2-0.5 m/s. I will then describe ongoing work regressing spectral residuals against housekeeping data — such as RVs, activity indicators, or instrument telemetry — under generative and discriminative frameworks. This regression establishes what structures in the residuals can be attributed to the effects traced by the given housekeeping data. These effects can then be removed to give more precise RVs.