Whether users access Hopper to travel from the Netherlands to Nova Scotia or from Cairo to Cancun, the travel mobile app's goal is simple: pairing its users with the most affordable airfare available. And, with trillions of data points at its fingertips (and billions more streaming in each day), Hopper is finding new ways to understand what users want. Leaders from Hopper’s engineering and data analytics departments spoke with us about the underlying tech and data that goes into booking travel, and what’s next on the company’s itinerary.
EMPLOYEES: 224 (45 locally)
WHAT THEY DO: Combining trillions of data points and powerful machine learning, the mobile app aims to help people save money and time in the flight-booking process.
WHERE THEY DO IT: Cambridge
IDEAL CANDIDATE: Those who are not only great at what they do but have the curiosity to constantly learn new things, challenge the status quo, and experiment — even if they fail a few times in the process.
Ken Pickering, VP of Engineering
Ken manages product, data and infrastructure engineering teams in Boston, Montreal, New York and Sofia, Bulgaria.
BEYOND WORK: Ken boxes five to six days a week at a local gym in and is training to, hopefully, compete at an amateur level in 2019.
You’ve worked in similar engineering roles prior to coming on at Hopper. How do those experiences compare to what you’re doing now?
This is really my first time at a company at Hopper’s stage, where we’ve proven ourselves as a startup, but the name of the game now is all scale and growth. The simple truth is that the opportunity to help run engineering at this phase of a company, with as much potential upside, is rare. I’ve loved a lot of my previous jobs, but this one is definitely the most ambitious for me.
Engineers are dedicated to certain “project lanes” at Hopper. What does that look like for an engineer, and how do you foster collaboration between engineers with those lanes in place?
I’m a huge fan of the cross-functional staffing model written on at length by Spotify, which is where the inception of these lanes sort of came from. Essentially, they give a team the ability to focus on real problems while not being pulled into a myriad of other development tasks. It also organizes your business around making throughput on specific things that are super important to growth.
Lanes make it easy to determine how we prioritize what we’re doing, but don’t remove the fact that we, as an engineering organization, still need to come together and produce it the best way we can.
The simple truth is that the opportunity to help run engineering at this phase of a company, with as much potential upside, is rare."
What opportunities are there to work on different products, features or tech stacks at Hopper?
There are so many potential opportunities to make progress at Hopper — I can easily staff someone in an area they want to make an impact in.
Guillaume and Shahab, who were core services developers, are now working on our new in-house platform for data science research and development support. There are a lot more cases like this at Hopper, but essentially, if there’s something that we’re working on that interests someone who works here, and they’re not doing it, chances are they could be soon if they let us know.
Ella Alkalay Schreiber, VP of Data Science
Ella leads the data science team and is responsible for driving home the data-forward culture that Hopper is built on.
BEYOND WORK: Since learning to scuba dive more than 10 years ago on an island in Honduras, Ella has spent nearly every vacation since in a diving site.
What kinds of problems or challenges are you solving in the travel and booking industry with data analytics?
The prediction algorithm upon which Hopper is based was engineered to take in massive amounts of data — around one trillion price points per month, with five years of historical data and many trillions of archived prices. It consistently learns based on feedback from user behavior to offer the most accurate predictions year after year.
Today, 25 percent of Hopper's bookings are the result of AI, meaning users are booking trips they didn't explicitly search for but that the app knew to suggest. Conversion rates on AI-based recommendation notifications are 2.6-times higher than ones for which the users explicitly searched. Hopper's unique environment — a closed ecosystem, one-on-one conversations with each user — has created an opportunity to utilize machine learning and AI to learn users' preferences on a much deeper level and help guide their purchasing decisions.
There are so many interesting problems to solve, so if you’re motivated by taking on new challenges, there are a lot of exciting opportunities here."
What does your team do to continually improve the experience for a Hopper user?
In 2016, we began testing a new recommendation algorithm built on top of our core platform, which sent users notifications about deals to alternate origins, destinations, months or weekends. With every conversion, it further strengthened the algorithm, thereby making future recommendations even more relevant.
These days, we are working on a new way to predict the scope of flexibility based on the intent users showed us in the app and a new model to recommend them alternative trips. A slower feedback loop, coupled with trillions of possible trips to recommend, makes training and refining the algorithm a much more complex task. We are currently running a bunch of experiments of the new modeling in the app and so far seeing a significant impact on user engagement, savings and conversions.
You were promoted from a data scientist manager to VP of the team in the past year. Does that speak to the kinds of opportunities for growth for a data scientist or engineer?
Absolutely — Hopper is really good at providing unlimited growth opportunities. There are so many interesting problems to solve, so if you’re motivated by taking on new challenges, there are a lot of exciting opportunities here. We’re in the process of scaling and building out team structures, so it’s definitely an exciting time to join.
Huan Lai, Director of Engineering
On top of hiring and scaling Hopper’s culture with the growth of the company, Huan manages the teams responsible for the back-end development that powers the app’s flights product, as well as the data infrastructure team that builds its data pipelines and model building platform.
BEYOND WORK: A new father, Huan spends his time away from work with his 9-month-old son and family, but he still maintains an interest in collecting strategy board games.
You worked with Amazon prior to coming to Hopper. What about Hopper made you want to make the switch?
Hopper has a very modern tech stack, and I was very excited to get more hands-on experience with some of the latest and greatest open-source technologies on the market. We currently use Swift for our iOS app; Java and Kotlin and ReactiveX for our Android app; a Scala microservices backend architecture on top of HBase and HDFS; and Kafka and Finatra and Python, Spark and TensorFlow for machine learning and business intelligence.
Explain how machine learning, data and AI play into the Hopper experience.
Our original claim to fame was our combination of machine learning algorithms and big data to accurately predict whether airfare prices are going to go up or down, with 95 percent accuracy up to one year in advance. But now we’re also using AI to understand the underlying intent of the user and recommend trips they weren’t even originally thinking of.
For example, a user that watches a bunch of trips in late December from Boston to Punta Cana, and Boston to San Juan, and Boston to the Bahamas is probably just looking to go somewhere warm during their Christmas vacation. So, we might send them a notification when tickets from Boston to the Virgin Islands dips below the trips they were originally looking at.
Hopper has a very modern tech stack, and I was very excited to get more hands-on experience with some of the latest and greatest open-source technologies on the market."
What are some exciting projects that your team is currently working on?
All of our machine learning services deployed to production thus far have been implemented as one-off services designed specifically for that use case, leading to months of development to get something out the door and a severe operational burden.
A major project we’re working on today is a single end-to-end platform allowing data scientists to train, deploy and evaluate machine learning models without having to write any application code. We’re actually rolling out the first phase of this project in December, allowing data scientists to deploy and serve versioned TensorFlow models in production with a simple API.