Avari Blog

Real-Time Predictive Content Delivery: An Interview with AVARI’s Senior Developers

[fa icon="calendar"] Feb 18, 2018 / by Natalye Childress


At AVARI, we’re excited about our recommendation engine, which relies upon predictive technology and open-time personalization to deliver dynamic content in emails. It’s fast, it’s lightweight, it’s broadly applicable, and it’s smart.

Personalization technologies have traditionally been heavy, expensive, and slow, taking months to achieve ROI. AVARI’s technology was born with a mission to drastically improve on these aspects and provide a better solution to the market.

To explain how we were able to do so, we chatted with VP of Engineering Christoph Bünte (CB) and Senior Software Developer Kacper Bielecki (KB), the brains behind our secret sauce at AVARI.

How does the technology work? 

CB: Our technology focuses on gathering data on browsing behavior. Most personalization technologies focus on customers making purchases and using transactional type of information to make judgments about subsequent purchases. We began by thinking, “With so much data available when a customer is shopping and searching through a website, why not use some of that to influence our content recommendations?” As a result, we built a lightweight javascript snippet that integrates to the frontend of a website, similar to Google Analytics. This tag is flexible and robust, and can be added to a container tag system like Google Tag Manager. We leverage some of the latest developments in web tracking by using browser fingerprinting to capture real-time visiting behavior including dwell time, clicks, basket items, and more.


What drove the decision to use web-tracking technology in creating the AVARI product?

CB: Previous personalization technologies were heavy, based on server-side integration, and expensive, both from a cost and resource perspective. Web-tracking technology has become very popular with services like Google Analytics, Optimizely, and others. Our decision was to keep it lightweight and flexible, allowing clients to be able to integrate it quickly, easily, and across all of their platforms.

There is always a concern that tracking this much data can have latency and scale issues. Is that a concern?

CB: Scaling is always a concern, but we’ve built our technology on some of the best open-source technologies that have experience in scale. Building on this enables us to own the software stack and make decisions as a platform that serves our customers effectively but still relies on the community for assistance.

What are the benefits of gathering all of this data for the recommendation engine and not simply relying on direct purchase history?

KB: Consumer purchase history tells some of the picture, but the journey customers take before they make the purchase is an area technology solutions have not explored before. Some signals from this data may not be of use initially, but over time, our models improve based on the variety of signals we receive. Just as Netflix continues to seek to improve its recommendations, we believe that the advances in personalization technology live within the consumer journey before making a purchase.

How can that information help make recommendations better?

CB: We get more of those events. What we track right now is if a person looks at a product, or if the person puts something into the cart. And we assign those actions with an engagement level and feed that to the recommendation system. As those events happen more often than an actual transaction, we collect more paths of those events for single individuals.

The information is anonymized, but we feed it to the recommender to figure out the type of user, as well as similar users who have been on the same page before and possibly took the same kind of actions, and we look at what they lead to. Based on these things, we can make recommendations to a person who’s new to the shop, or even an unknown. This varies from most recommendation systems, which have the problem where they cannot recommend something to a totally unknown individual.

You mean the “cold start” problem.

KB: Yes. Of course, most information about users and customers comes from the transactions, at least initially. But we also have the information about the content that people looked at and didn’t buy or download. This is the real benefit. 

Additionally, if there are 10 times more people who just browse the website as opposed to buying, we also have the possibility to profile the user who is only browsing the website. In this case, it won't default to only recommending top products. Instead, we'll always have some information to start with. So while I think the behavior of users is reflected more by transactions than by going through the shop, I also think browsing behavior will allow us to, in many more cases, deliver accurate recommendations to people who otherwise would have no recommendations.

Let’s talk about the decision to use open-time technology which is used to deliver the predictive content. How is that being thought about and built?

CB: Essentially, the open-time technology inserts a dynamic block as an image. You can think of it as a blank canvas, enabling our technology to display anything to the customer at that moment. Whenever you open an email, that image is fetched from our server, and with that image comes a unique identifier, so we know who is opening the email. We then immediately look at our recommendation platform to deliver the most relevant, predictive content in that moment. 

This is important because our engine will do this calculation right as the customer opens the email, which means if there are products that are performing well or driving great revenue in the time the email was sent, we can capture that and recommend it, ultimately leading to better performance.

How was it being done before? What are the benefits of the new way of building it, versus the legacy of RetentionGrid?

CB: Early on, we rendered the HTML snippet on the fly, we cached it for five minutes or so, it turned into an image, and then it was delivered. And it was fast enough, but it all depends on so many things. Like how fast can we deliver? How fast can the shop deliver that product image? And for things like live social feeds that becomes even more important.

There are several things that set AVARI apart from other technologies in the marketplace, and speed is one of them. With the ability to inject dynamic content into emails in less than 100ms from the time of open, we’re using real time to craft a unique customer experience in the blink of an eye.

To learn about AVARI technology in more detail, check out our webpage about it.

Topics: Developer Thinking

Natalye Childress

Written by Natalye Childress

Natalye is the Writer and Editor at AVARI, where she’s in charge of crafting copy, creating content, and fine-tuning on a sentence level. When she’s not reading or writing, her interests revolve around road bikes, craft beer, vegan food, and live music.