Could you specify what kind of data selection and pre-processing was required and why exactly it was hard to implement it in a real-life setup?