Over the last two years we have developed a Neural IR model for complementing the traditional search system of a large fashion e-commerce company. The ML model has been rolled out in 16 countries and is used for increasing recall of low-result queries as well as for query-dependent ranking powering 20% of traffic today. In this talk we will present the different stages of the model development, the feature representations we have chosen, how we generate massive amounts of training data, and how we manage the tradeoff between learning from big data and staying efficient. We will demonstrate the use cases and present successful results from offline and online testing.