Chief technology officer (CTO) and Co-founder at Picnic, Daniel Gebler, has been leading the technical team since the online supermarket started in 2015. At the AI Founders Night by Slimmer AI and ai.nl, he talked about different challenges and solutions, like changes in customers’ search behaviour, safe driving from their delivery drivers and the change to an automated warehouse.
“When we started, we had to figure out the consumer position, which is quite simple, deliver all food the next day and offer the lowest price and free delivery. The reason why this is important is that this is the additional implementation of disruptive innovation. If you build something better than the offline variant, you open a market that was unavailable for customers. In 2015 only a minimal amount of people used online shopping for food. When we did our research, we didn’t look at people already shopping online for food, but we looked at the group who had never done this before. All these customers didn’t feel comfortable yet doing this. The key thing is that if you don’t ask for delivery fees and can make this sustainable, you have a business that will stay around for a long time.”
“Due to COVID-19, there was a huge shift in demand from offline to online. Besides this demand that has exploded, the entire consumer behaviour has changed in a tormented way. Larger orders, different types of day, but the searching behaviour has changed the most. Before COVID-19, people searched for chicken and milk, but in March 2020, people started to search for toilet paper and painkillers. The only thing you can do now is start to train your model again to get used to this shift in search behaviour from the customers. This new model had to be trained on a minimal dataset, so we had loads of fun trying to make it work.”
From a challenge to a solution
“We’re constantly picking from a challenge to a solution. If there’s no consumer or logistic problem to solve, we won’t apply any AI. Of all the different challenges, there is interesting one. We’re trying to make shopping as easy as possible. If you go to a site like Coolblue, you’ll be shopping for 20-25 minutes and buy around two or three items. The average amount of items you buy at a supermarket is around 30. What we see is the kind of frequency, so the rebuying is also high. You never buy the same products again and again at Coolblue, but you buy certain types of food every week, around 50% of the products you buy in a supermarket are the same as last week.
“The 50% that you rebuy is straightforward, because you check what you bought last time, and you repurchase it. However, this is also a bit of a challenge because you don’t know which 50% this will be. The remaining 50% new are tough to predict if you solve that. Much harder than non-food because there is no clear projection from the last week to what you will buy this week.”
“If you want to decrease shopping time from 30 to 3 minutes, you need to think about building recommendation systems that go beyond the basic recommendation. A typical recommendation system, like what item to buy next, is related to the current item. If you want to go beyond this, you simply want to recommend not one but multiple items. If you want to change this, you don’t need to investigate recommendation systems that not only recommend one item but something like four or six. Then you end up in an interesting challenge. The challenge is: if you have a currency of a simple item that you want to buy, if there is no correlation between those different recommended items if you then recommend twelve items, the likelihood that you buy twelve items in one go is less than 30. We looked at many different recommendation systems, built our new model and fed it all past purchases of consumers, and you get the most likely item to buy the next time you go shopping. However, you now have additional challenges. You have all kinds of exceptionally balanced models. If you have a summer break, your behaviour will differ from that in the spring. You also need to model this behaviour.”
“We have around 3000 little Picnic cars on the road across The Netherlands, Germany, and France, and we need to ensure safe driving. We collect a lot of data from the streets of driving our vehicles. We identify what is safe driving and what is not safe driving. However, we don’t keep this data ourselves; we give this back to our drivers. After the ride, you’ll get a notification where you can see what your driverscore is. The driverscore is a score between 0 and 100; if you get over 80, you have driven safe. If you get a score below 80, you’ll get recommendations on how you can improve this and why your score was below 80. If everybody inside a hub, a city, drives safely, everybody will get a bonus to encourage the drivers to all drive safely. The next step is not only the driverhub, the tool for the drivers to check how safe they’ve driven, but the servicecoach. This is the recommendation for interacting with a customer at the door. We can recommend the best interaction with a customer; for example, somebody had a birthday recently, so you get a recommendation based on this type of event. This is also the other way around, and maybe the last order didn’t go well. Then you get a recommendation on handling the next order differently and apologize, for example.”
The warehouse automation project
“We have warehouses of around 20.000 square meters where order picking and collection are implemented. What happens here is that you have a lot of shelves, and in the morning, you’ll get around a million products that are being delivered to this warehouse, and you need to put these products on the shelves. Around noon these products will get picked up based on the orders the customers have placed. This picking works quite well, but it’s not the most efficient way. Instead of you going to the product, the product will come to you. This is an obvious way of automation. We changed this, and now we have warehouses of around 50.000 square meters that bring all kinds of products to the order pickers. The planning problem that we need to solve is how do you route a product from a storage location without creating congestion. This also brings a timing challenge. A product destined for consumer X must arrive at the same time as the other product, which is also for consumer X. We tried to solve this by simulations. All the active machine learning models during the day need to be replaced by night on an alternative configuration, and then we get the answer on how the day would’ve run if this model had been slightly different.”
“Lastly, I would like to give a few lessons to learn from founders and engineers:
- Dream big, act small
- Mission first, data as support
- Data science first, AI second
- Launch first, scale second
- Great products come from small teams.”