Delhivery

I worked at Delhivery from October 2016 to October 2017 on three projects as a part of a team to devise data-driven solutions to business and optimisation problems in the logistics industry.

RTO (Return to Origin) shipments are those which are not accepted by the customer at the time of delivery and are required to be returned to their original pickup location. The task was to accurately identify for a client when a shipment has a high probability of return (i.e. of not being accepted by the customer).

• Cleaned & pre-processed data, and extracted features from raw data to create combinations of features from customer behaviour, shipment attributes and client characteristics.
• Built a decision tree regression model using both categorical and continuous features that can accurately predict the probability of RTO based on the historical data available for that customer.

First Mile Logistics involves the movement of products/shipments from a retailer (client) to a Processing Centre. In the project, the task was to predict the volume of shipments to be picked up from a client warehouse on a given day in the future to aid in the planning of an optimised vehicle route for the first mile pickups.

• My work on the project involved predicting the number of packages to be picked up from a client warehouse, along with physical volume of each package picked up.
• Wrote code in Python to clean and pre-process the dataset containing the details of all past pickups, and segment it on the basis of similarly behaving client warehouses.
• Built a model for each segment, taking into account a different feature set for each, to accurately predict the number of packages and the typical volume of each package.

Hyper-local deliveries (HLD) involve picking up items from a merchant and delivering the item to a customer, while ensuring an optimum turn-around time (TAT).

• The first part of the project focused on predicting the TAT from statistical distributions, given the time of day and distance to be travelled along with other parameters. This involved clustering consecutive hours based on similar order-arrival and traffic conditions to accurately provide an estimate of TAT for each cluster.
• The second part of the project required predicting the number of delivery staff required during each hour of the day, so that the average TAT during the day as well as the average TAT for each hour were within a given time range. The approach involved first using a deterministic model to calculate an initial guess for the number of delivery staff required and subsequently using that initial guess to determine the optimum number for each hour.

• Presented knowledge-sharing sessions for colleagues on tree-based modelling and ensemble methods