As part of a four-person consulting team of VTEX software engineers and outside data scientists, I used neural networks with adapted NLP techniques to create a recommendation system trained only with previously-purchased carts and no sensitive user data.

Its universal nature, making no assumptions about what the items are or how they're expected to relate, means that it can work in any of the platform's over 2500 stores. It also managed to outperform baselines by 350-500% in its initial version. We also built a fully functional web application to showcase examples of given recommendations in practice. Feel free to watch the presentation below, which goes into more detail.

Technologies used: Python (pandas, matplotlib, scikit-learn, keras, tensorflow), SQL