The start of new things can bring new perspectives, and new perspectives can bring great things and new ideas which we of course always value at Essent. In this article, we will be talking about the start of the Databricks journey at Essent for Mark-Danney Oonk and Ruud Goorden.

A bit of a backstory on the models

¨Ruud and I have been working on migrating some of our data science models. These models originate from data from our NPS monitors. At each of the touchpoints and journeys we have at Essent, we have what we call an NPS monitor. In the NPS monitors, we measure customer satisfaction and if customers would recommend Essent as an energy provider on a metric scale between zero to ten. This metric is called the NPS (Net Promotor Score).

Furthermore, we ask them if they have any additional comments to please fill in a text form. Instead of having our people manually go through those texts to understand what customers are talking about, we trained models to recognize the text and label it so that we can make a report. With this, we won't need to spend a ton of hours labeling the text manually. ¨- Mark-Danney Oonk

¨The models automatically assign topics to the textual feedback that customers give to us. Since the textual feedback is too hard to analyze because it is unstructured data, we make it more accessible by assigning certain topics. These topics were determined initially by our business stakeholders. If the customer says ´Hey I was very satisfied, but I couldn't find the bill properly,´ from this you can get information about the findability which is an aspect of the bill. Then the model can automatically assign it to findability which is a topic that the customer is talking about, and in that way, we can easily analyze these topics. With this, we can dive into seeing which type of customers say I can't find it very well or how often the findability is an issue for customers so that we can work on that with our process teams. So that's what the model does, automatically assigns topics so we can analyze them better and take appropriate action. ¨- Ruud Goorden

Can you tell us a bit more about the monitor?

¨At each of the important touchpoints and journeys we have at Essent, we have what we call an NPS monitor. Basically, it is a rating between zero to ten. So, if you give a rating above nine you are considered a promoter, seven to eight means that you are neutral which means that you just like the product but also means that you aren't super into the product either, and anything lower means that you weren't happy with some part of your experience and you are called a detractor.

These scores get taken together and then we get a final score (NPS) from that. With this, we can see as a company how we are doing in relation to our customers and how satisfied they are. ¨- Mark-Danney

How long have you been using these models?

¨We have different monitors and we started with the invoice monitor approximately a year ago. With the invoice we first made the model then other people in the company got interested. Models for our new customers (I Join) soon followed. At the moment we are doing the development of the digital monitor and contact monitor. Basically, the framework that we are now building is structured in such a way that we can quite easily add the new monitors if we have data for that. ¨ - Ruud

Did the creation of the models take long?

¨The modeling process itself doesn't take long but the process before the modeling which includes manually labeling takes some time. This is because you must have discussions on what type of topic it should be. For example, someone might say it should be topic A while another would say it should be topic B. You need to have these sorts of discussions on what type of topics they are and what is the definition of those topics. With the modeling, I would say that we have quite some experience now and we can see quite fast which models work well. ¨- Ruud

What is the business benefit from this?

¨The business benefit is around being able to access the unstructured data and analyze it. This is its core benefit, and it takes away a lot of manual work like going through all those texts and seeing what it is about. Because they use to do some manual labeling like every month and doing a count of what type of topics there are and now it just it automatically and saves us time. ¨- Ruud

What is Databricks?

¨We have these models that Ruud made about a year ago and he was executing them monthly manually. This is of course not a sustainable way of productionalizing our models, so we decided to migrate his models to Databricks. Databricks is a data science platform where we can host our models, and our code and we can spin up clusters of very powerful virtual machines that can then run those models and the code that we need to prepare and analyze the data. Lastly, Databricks will offer our data scientists a good platform for cooperative exploration and experimentation for future models.¨- Mark-Danney

What sort of problems did you encounter along the way?

¨Databricks was quite new to us at Essent. So this project was a nice way to uncover issues regarding setup and infrastructure we hadn’t thought of before. Moreover, our team had already made some utility functionality on top of Databricks to improve user-friendliness, which we could now test and improve incrementally. So our teething problems will help the rest of our team with the following projects.¨- Mark-Danney

Lastly, what went well?

¨The main thing that went well was the cooperation between Ruud and me. From day one we were both very eager to learn this new platform. From his side, the engineering was something new, and for me the data science aspect, but we found a great overlap which was really cool

And of course, our whole team with the engineers that built the framework on top of Databricks, were very enthusiastic for someone to finally use their codes because it had at this point not been tested yet¨- Mark-Danney

This is just the start of the future of Databricks at Essent and we can't wait to share more about it with you. So until then see you on the next blog post!