Differentiated and innovative online gaming with data + AI

Data infrastructure stifles collaboration and innovation

In a rapidly growing industry, where competitors offer the same live matches, the same betting opportunities and the same casino games, Kaizen, the fastest-growing game-tech company in Europe, knows that data is a differentiator. “We use our data to improve our products and service levels. Responding to customer behavior insights builds customer loyalty and revenue,” said Constantinos Liapis, Head of Applied AI at Kaizen Gaming.

However, Kaizen’s first personalization project began with the ambition to enhance the customer journey. A typical Saturday might offer 300 live sporting events, and it was taking too long for players to search the list and find events of interest, so Kaizen wanted to offer players a top 10. “We had already started a training model, but it was running on local PCs, which just couldn’t cope with the volumes of data and wouldn’t scale. We had to use just six days of data – not enough to generate meaningful personalization of the sportsbook. Plus, the lack of collaborative functionality was holding innovation back. We couldn’t work together as a team.”

The team also wanted to increase customer retention by automating the manual bonus and reward process, “CRM systems could take two weeks to award a customer and, by this time, the customer was disengaged. They needed to be real-time to be relevant.”

Kaizen has huge volumes of data, especially in the casino, but access and processing limitations meant they couldn’t produce a granular enough view. They were missing out on valuable customer insights. “We could see activity per day, but we needed to know which customers play which games, down to what spins minute by minute with the previous infrastructure going to minute- level details was simply not possible,” said Liapis.

Upping the ante with Azure Databricks

Kaizen has dramatically upped the ante when it comes to dealing with large data sets at speed. “We’re greedy for building new things, and now we’re on a cloud model we won’t run out of resources.” Creating a bigger cluster is a matter of clicks away and different clusters can be set up, each suited to different workflow tasks, to maximize resource efficiency. “We group jobs that have similar resource demands together. For example, for model training RAM is a priority; while for performing predictions it’s CPU.”

And now, at last, Kaizen teams can work together. The use of a common space for all components, models, notebooks, experiments and data allows for better coordination and instant reference. ML engineers can work closely with big-data architects to build streaming functionality, calculate features, store them in Delta Lake, create the feature vector and push to the prediction stage. “Databricks allows us to view results and work on the same code concurrently,” said Liapis.

Azure Databricks has transformed machine learning development and deployment at Kaizen. They now use MLflow for ML development, training models, data analysis and model deployment, experiment tracking and model registry. “Now we can see model metrics during experimentation. We can keep track of different experiment settings, and the team can easily compare different approaches coming from any step of the flow. It has improved our model deployment process significantly,” said Liapis.

Winning big by delivering better insights and revenue

For Kaizen, Databricks means more data and more detail. Creating data sets is fast and simple: What had taken 240 hours, now takes just 20. “In the past we waited for 10 days for the data set creation to complete. Now, with a similar case, that number is reduced to just a few hours,” said Liapis. We can query our data, we have on-demand experimentation and training. We can do two or three experiments per day, using big clusters, and we can combine ML with the application layer to deliver an end-to-end product. We would simply not be able to do that without the Databricks platform.”

All legacy use cases will now move onto the Databricks platform including projects from early days of ML experimentation, “The porting functionality is incredible,” said Liapis. “The sportsbook personalization use case was moved to Databricks in just four days. I expected a month, but suddenly in just a few days we had 365 days of data to train with.” Kaizen now uses Databricks to generate predictions and propose the top 10 events on a daily basis. In testing, predictions and recommendations were compared with actual play and they hit 75%–85% precision across the top 100 events. “That’s a very impressive number for a recommender. We have made things easier and more relevant for the customer.”

Kaizen now understands interactions by customer segments and time of day, which means they can be more effective with advertising and bonus campaigns and promote offers that are of maximum interest. “We are close to delivering real-time bonus award payments,” said Liapis.

“Databricks has transformed productivity at Kaizen. We can do more ML and produce better results, faster. It’s solved data infrastructure-related problems and brought value to the business much faster than before. Azure Databricks is a massive acceleration factor.”