receive a plurality of damage evidence files comprising images of a vehicle damaged during an adverse incident;

identify vehicle information based on a vehicle identification number (VIN) associated with the vehicle, the vehicle information specifying at least a body type;

identify a damaged panel of the vehicle based on the damage evidence files by applying the damage evidence files as input to a trained deep neural network (DNN) that has synaptic weights that have been trained using semi-supervised machine learning techniques to encode historical correspondences between damage evidence files and vehicle panels, wherein the DNN comprises multiple stacked layers including an input layer, an output layer, and multiple hidden processing layers between the input layer and the output layer that perform transformations for aspects of automated damage appraisal, and wherein responsive to the input, the DNN identifies the damaged panel;

determine repair information related to repairing the damaged panel based on the vehicle information;

determine at least one part (i) that is adjacent to the damaged panel and that may be damaged during repair of the damaged panel based on the repair information and at least one rule of adjacency;

identify at least one repair estimate fragment for the at least one part; and

generate a repair estimate line comprising the at least one repair estimate fragment.