Fast delivery platform
for high quality prepared food
As part of the existing Azbuka Vkusa business, it was necessary to launch a new startup to capture the segment of fast delivery of ready-made high-quality food. Azbuka Vkusa had about 200 offline points with a ready-made assortment. We had to create a service that would allow users to order ready-to-eat food delivery from these locations within a guaranteed time (30-60 minutes) in certain delivery zones.
In the project, I performed a multifunctional role - CTO, product manager and data analyst. Managed a team of 30 engineers to ensure product development and launch.
- Development of a commercial platform (catalogue, receipt calculation, shopping cart, users).
- Creation of a product website.
- API development for mobile applications.
- Integration with courier services operating in delivery areas.
- Integration of the platform with internal accounting systems (accounting, CRM, WMS).
Story 1: Load balancing for couriers
Uneven workload of couriers: at one time they are idle, at another they are all busy, and there is no one to process new orders. This led to delivery delays and customer dissatisfaction.
- User Clustering: We analyzed the order history and users of large clusters with similar preferences. The catalog was ranked according to the taste preferences of each cluster. For example, some users preferred to order seafood dishes in the evening, while others preferred meatless dishes. Each dish on our platform has its own components that affect the value depending on and create personalized recommendations.
- Model Testing: We developed a prototype model and tested it on 5% of website users and in local apps. The checkout speed increased by 20%. After that, I handed the models over to engineers for rolling out to production and retrained the models daily so that users applied the best recommendations every day.
- Real-time Personalization for Guests: For users ordering as guests, a fast K-means-based model was developed and worked on for the first time. This model ranked the catalog according to user preferences without requiring retraining. After authorization or registration, the user history was exported to the main, more complex model.
Story 2: Catalog Personalization
We needed to increase the speed of online checkout so that users could find dishes they would like faster.
- User Clustering: We analyzed order history and created clusters of users with similar preferences. The catalog was ranked according to the taste preferences of each cluster. For example, some users preferred to order seafood dishes in the evening, while others preferred meatless dishes. Each dish on our platform had its own ingredients, which allowed us to identify dependencies and create personalized recommendations.
- Model Testing: We developed a prototype model and tested it on 5% of web and mobile app users. The checkout speed increased by 20%. After that, I handed the model over to engineers for production and ensured daily retraining of the model so that users received better recommendations every day.
- Real-time personalization for guests: For users ordering as guests, a fast K-means model was developed that works in real time. This model ranked the catalog according to user preferences without requiring retraining. After authorization or registration, the user history was exported to the main, more complex model.