Jordi Esteve

Glovo is a food delivery company. My team was responsible for ranking restaurants in the restaurants wall (aka feed). It was a key app component because all traffic was going thru it. The limitation was that the ranking was done using a conversion rate (cvr) model that precomputed the cvr for each store every day. This has several limitations: - A store would have a fix cvr each day - Not personalized to the user - Not scalable Given this, I proposed to do it in real time and personalized. The solution is: - Prefilter stores for a user: using store availability and distance to user - Given user features - Given store features - Compute cvr score for each pair To achieve this there are several steps: 1) Get the data. I designed and built a pipeline that generates the fact table of user-store logs, so we can train a model. I designed and built the pipeline for dimensions, i.e user and store features. These run daily and are both used to train the model and features for real time serving 2) Train a cvr model. For that, I used lightgbm + calibration layer. Lightgbm was the best trade off between efficency and performance 3) Build serving layer. The serving layer was a FastAPI api with the model, and two SQL databases, one with store features and the other with user features Results: - Increased revenue by 12M/year!