Step (1): Get access to frontal photographs of a large random sample of borrowers along with their financial history and records of any defaults.
Step (2): Extract a vector of facial features from each photo. Similar to the feature vectors used to judge attractiveness in the work of Cohen-Or et al. http://www.cs.tau.ac.il/~dcor/
Step (3): Have users apply for a loan by submitting a mugshot. Use support vector regression or some other similar technique to predict the probability of default.
Step (4): Profit