Artur and Tomasz tackled a problem of finding cross-selling opportunities in a B2B setting and extrapolating them to new markets. Application of Deep Learning methods on customer-specific time series of historical purchases led to promising recommendations validated later with help of active participation from the sales team. A short time-frame for the process of building models did not allow live validation of results, while it is not possible to fully validate opportunities on just historical data. To reduce this problem they carefully designed a scenario-specific loss function to avoid penalizing model on plausible recommendations.