Unit economics optimization for e-grocery company

How Predictive Analytics helped optimize the number of pickers' shifts. 

We collected the data on client's competitors and defined pickers management as the main source of improvement.

Our solutions were targeted at 2 goals.

First, we wanted to better predict the amount of work (basically, the amount of orders) that was required at a given hour. Second, we tried to make the work more efficient. First target was measured in standard bias and variance terms. Specifically, we picked MAE and WAPE metrics for that.

Second target was measured as an average time to collect the order.

To better predict the amount of orders we changed the existing approach in spreadsheets to more advanced Machine Learning based model. This allowed us to account for special occasions like holidays and to faster react to changed conditions. After many iterations our client were able to predict the amount of orders 20% better than it was done before. For the second goal we worked closely with pickers and asked their vision on how to help them become more efficient. We found that the map of the shopping room helped less experienced pickers tremendously so we built it based on the pickers logs. For more experienced pickers we found that optimization algorithms can only help with large orders that have sophisticated optimal routes. Another solution was to batch the orders for 1 picker to collect them in parallel. All of the above allowed us to decrease the time required to collect an order by 25%.

After testing our POC we implemented a model as a web application that builds routes for incoming orders. We deployed the app into the cloud-based kubernetes cluster so the client is paying only for the real orders that are routed. 

This project was one of the most challening for Tez Data consultants and took 3 quarters. We and our client learned a lot while going through this journey.