How Assurance Unlocked More Business Value with Starburst

By leveraging Starburst, Assurance was able to improve conversion rates, reduce costs, and enable robust modeling. Read the full case study here.

At the Datanova for Data Scientists, I had the pleasure of giving a Lightning Talk on the impact of Starburst and Trino on accelerating our time-to-insights, improving customer conversion rates, and enabling the creation of robust modelling for insurance recommendations. Starburst provided a single source of access for all our data and gave us the ability to see data real time, which is essential for our customers trying to enroll in insurance programs.

Assurance is an online distribution platform for insurance and financial products. It operates between consumers who are looking for coverage, the agents that have the expertise to educate on the available products, and the actual interface into the many products available. However, the world of insurance has many complex rules and regulations, so Assurance makes it easy for customers to navigate these complexities, and data gives us the ability to do this. So,  as customers progress through the buying process, we are running models through each step in order to give the most accurate recommendations. However, this was not always possible.

Before Trino, we struggled to join data from one database to another. There was a lot of copying involved which led to longer time-to-insights. In addition, this lack of efficiency made it difficult to scale to the needs of the business, data scientists were under utilized, and stakeholders couldn’t understand the delay for insights. This all changed once we implemented Starburst, the enterprise grade edition of open source project, Trino. By having a single point of access for all of our data, we were able to unlock a lot of business value.

For my Datanova presentation, I focused on one aspect of our insurance offerings: Medicare Advantage plans. Medicare is a federal health insurance program run by the government particularly for those who are 65 years and older or for eligible individuals  with disabilities or advanced diseases. Medicare Advantage plans are health plans which are offered and administered by private sector insurers. Medicare Advantage plans adhere to all of the rules set by Medicare. Private sector insurers are able to offer better economies of scale and the government can pay the insurance carriers directly. There is a time period each year called the Medicare Annual Enrollment Period (AEP), which is a 7 week period during which an eligible person can sign up for, change, or disenroll in their Medicare plan. It is a very complicated process and depends on a variety of factors; therefore, shoppers need help to navigate the complexities, which is where Assurance comes in.

Preparing to Launch: Optimizing our Agent Licensure Strategy

In order to enroll a shopper in a Medicare Advantage plan, agents must be licensed by the state the shopper resides in and the agents have to be appointed by at least one carrier in the state. Therefore, a business problem arises: How many state license and carrier appointments should we buy for each agent to best meet the needs of our customer base and maximize our profit over the annual election period?

The solution is a mixed integer optimization model which allows us to define our constraints, feed in data, iterate over many scenarios, output those scenarios, and analyze the outputs for optimal solutions. Data inputs included the cost of state licenses and carrier appointments, the varying level of skill that agents have, the various times that agents work, and the variation in shopper volume based on time of day or time of year.

Before we implemented Starburst, we had this data in many different sources, and there was a lot of friction when trying to aggregate all of the data available. However, with the implementation, data scientists don’t have to deal with the complexities of aggregating data anymore. Instead, they can focus on analysis and model formulation. The business outcomes have also been greatly affected. There was accelerated time-to-insights, which improved user experience, and it allowed for robust model development, which increased revenue and reduced costs.

When Logging Just Won’t Cut It

I’m sure you’ll agree that real time data is essential for any technology business, so, if basic functions aren’t working, businesses can’t adjust the course based on consumer behaviors at the time. So, when our logging tool could not handle the amount of requests from our employees, we decided to replace it with a data lake with a simple five step process: we inspected the logging indexes to identify key data sources, then, we cross checked it against what was available in the Trino catalog of connections, we identified data gaps, subscribed to streaming events for those gaps and wrote them in S3, and, finally, we connected Tableau to Trino- all of this was powered by Starburst. What resulted was real time analytics and the ability to build an insightful dashboard for our operators.

After this move, we massively accelerated our time-to-insights and allowed for new data to create visualizations. In addition, we were even able to present this dashboard at a company all-hands the day it was installed. It is still used daily and is very low-maintenance.

Rocking Recommendations

In order to determine the right product for customers, we needed access to the right data and a rapid development process. Therefore, the ability to rapidly test, improve, and implement product recommendation models was essential, and today Trino, the SQL-based MPP query engine is at the center of this. Through Trino and Starburst, we were able to optimize the agent licensure mix, get visibility back into operations at critical moments, better serve our customers, and increase our revenue. We are now concentrating on scaling up the Lakehouse architecture in order to better focus on analysis and business value.