Kreditech

I went for a micro services approach as this allowed me to quickly swap out services if needed. We also performed split A/B testing to understand which models performed best. The machine learning models themselves are implemented in R and deployed using RServe. Other services are built in Scala, mostly using the Akka framework.

• R/RServe
• Scala (Akka stack)
• Docker & Kubernetes

People are more important than processes. In contrast to many Agile environments we adopted a very lightweight process. We held one team meeting per week, we didn’t follow Scrum or Kanban, and we didn’t write user stories.

Conventional wisdom would say this is a recipe for a disaster, but we were successful. Why? because we had a very close relationship between the product managers and technical teams. I fully understood the direction the business was heading in and was able to make technical decisions to support this.

We also adopted a federated team structure. In essence, the sales team had a group of developers to support them; the risk team had the same. The platform as a whole worked well because each team operated independently and offered a “service” to other business units. The teams were highly focussed and understood the business goals and technical implementations well.

This structure also allowed the business to scale up quickly. When we wanted to introduce a new product we could pick from a range of in house services. We had confidence that the service we were using were solid, and the people behind them understood them well.