In March, we announced Transaction Risk API, our latest solution to fight payment fraud and improve the efficiency of authorizations, built specifically for models. In this three part blog series we’re diving into why we built it, how we built it, and what value it brings to the market.
Today, let’s talk about the how.
Fraud is getting more sophisticated as the payments industry continues to see massive disruptions, and merchants, lenders, and payment processors are turning to machine learning models to detect suspicious qualities in a transaction. Our goal with Transaction Risk API was to provide a powerful tool to help in this space.
One of the biggest factors is time. Customers expect lightning fast speeds when it comes to online transactions, so we needed to create a low-latency product that delivered high impact data insights at incredible speeds.
The result? 99% of queries served in under 100ms.
Transaction Risk API represents a tenfold reduction in response time over other solutions. Let’s look behind the curtain to see the months of team effort that made this possible.
The key to speed with Transaction Risk API is engineering innovation. The Ekata team made improvements in three main categories:
Synchronizing downstream calls: Depending on the query parameters, up to 100+ downstream calls are made to form an API response. We analyzed these staggered calls, prioritized them, and made them as synchronous as possible to reduce wait time. Now, all primary validations are initiated at the same time, then secondary calls are made, and so on.
Reducing overhead: Transaction Risk API was still plagued by a mysterious application overhead. To diagnose it, we extensively profiled all of our applications to gather data and identify bottlenecks, then built out dashboards to track the progressive improvements. By deconstructing single queries and then incrementally adding complexity to the request, we were able to analyze impact and then experiment with different deployment settings and environments. We also improved the efficacy of our Identity Graph calls to find optimal balance between latency and coverage.
Ensuring peak performance: The Ekata team explored several options to ensure connections between the application and downstream microservices performed at peak. This included increasing keep-alive timeout on all the internal microservices, testing with an HTTP2 upgrade, and digging through countless TCP packets. We mean countless packets. Talk about getting real.
Data Provider Partnerships
One of our biggest focuses as a company is sourcing quality data. Our products, our customers, and our reputation rely on it. For Transaction Risk API, we ensured the best coverage results through data provider partnerships, making live external downstream API call outs in certain cases.
Our team worked directly with these external data providers to reduce their response times — in some cases up to 5x. We did this through strategies like creating closer access to their server (thereby reducing network latency) or expanding cache coverage and values.
Advanced Network and Machine Learning Signals
Our data scientists also leveraged our machine learning expertise to build a number of proprietary attributes that had historically been third-party call-outs. Bringing these attributes “in house” contributes to a better product because it allows us to respond to market feedback more quickly, improving predictability. Not to mention immense reduction in response time.
To make an even stronger product, we also crunched the data to discover the optimal number of links to fetch from our identity graph (i.e., address-to person, phone-to-person, etc.) in order to balance ROI against response time.
At every step of the way, we maintained and continuously re-ran representative samples to assess the impact to coverage and AUC (area under curve) metrics from new developments.
Transaction Risk API: Intelligently built for performance
As a market-driven company, Ekata continually works to improve our products based on our customers’ needs and feedback. The Transaction Risk API was engineered to be easily integrated with existing solutions in order to provide the lightning-fast response times you need to fight payment fraud.
It provides low-friction identity verification data to verify good customers while segmenting questionable transactions into a higher-friction path by scoring the overall risk of an identity using email, IP, phone, name, and address. All in under 100 ms.
Transaction Risk API is also powered by the fastest and most reliable tech stack in market, allowing it to scale to any low-latency, high volume model requirements. For more information about Transaction Risk API, visit the product page or request a copy of our Transaction Risk for machine learning user guide.