Carson Leung

During Trulioo’s hypergrowth phase, the engineering and customer success teams had issues keeping up with the 400+ data vendors worldwide. Each vendor had its unique maintenance schedule, downtimes, and other disruptions. The company could only react to performance degradation once it had impacted customers for a prolonged period. Often, this resulted in thousands of failed KYC and tens of thousands of dollars in lost revenue from manual onboarding.

In response, I spearheaded the design, development, and maintenance of an early warning system that proactively scans all data vendors and key clients for disruption in the company’s flagship products. The system immediately notifies the customer success, engineering, and management teams with a call to action if anomalies are detected. By automatically switching affected data sources to backup data vendors, this system prevented disruptions and minimized the impact on customers. This system made our team proactive instead of reactive, which was pivotal in securing larger customer contracts and expansion.

In Jumio, the business unit I integrated was responsible for developing the Jumio transaction monitoring system and behaved like a startup within a corporation. Its database at the time was optimized for quick writing but was very suboptimal for reading and updating records. Because of this, there was a lack of business intelligence infrastructure and data consistency, which became a clear blocker for any platform-wide data analysis and building predictive models.

I designed the architecture for a high-performance, serverless database that seamlessly integrated with the existing AWS components. Because the database was serverless, it was also extremely low cost. With this database, we developed ETL pipelines for further data refinement, which enabled business intelligence dashboarding and improved data quality above the threshold for predictive modeling.

Another business-critical feature that relied on the database was the anti-money laundering incident recordkeeping, which provided metanalysis and quick fetching of customers’ potential fraud incidents and enabled statistics to be reported upstream to regulatory bodies.

Mastercard had a very advanced biometric model that could derive and analyze over 400 user web behavior characteristics. Data points included device settings, geolocation, and the unique way a person types on a keyboard. Everything was combined to create a user’s fingerprint, then helped determine whether a user was exhibiting a known or unknown behavior. If a user showed too abnormal behavior, its traffic would be halted via an impossible captcha. The problem was there could be multiple users within an account. If not distinguished, the behavior of numerous users would be grouped, thus significantly reducing the model’s ability to tell if a behavior is abnormal.

I worked with data science and data engineering to develop a new rule set that allowed the model to recognize and segment traffic into primary, secondary, and unknown user behaviors within one account. This reduced false-positive rates in many of the company’s key accounts, sometimes upward to 12%. It was eventually adopted into a primary fraud detection method in its eCommerce marketplace clients.