You are building a model for a digital creative workflow product to predict whether a newly uploaded design asset will require significant manual revision before approval. You have been given a new tabular dataset collected from asset metadata, editing history, and reviewer feedback, and the main challenge is deciding how to turn the raw fields into useful model inputs.
How would you approach feature engineering for this dataset so that the resulting model generalizes well and remains practical to maintain in production? Explain how you would decide which transformations, encodings, and derived features are worth keeping as you train and evaluate the model.