Arpeely: Scaling an innovative data science platform globally with a small local team
About Arpeely
Data science startup Arpeely uses machine learning and feature engineering to discover hidden opportunities in online advertising. It processes dozens of billions of predictions daily and cherry-picks traffic based on in-app or post-conversion behavior KPIs.
Tell us your challenge. We're here to help.
Contact usAbout DoiT International
DoiT International is a global tech consultancy that helps businesses build and scale cloud solutions.
Arpeely launches and scales its innovative ad-tech platform on Google Cloud, managing global ad operations with a small team by leveraging compute, data, and machine learning Google Cloud solutions.
Google Cloud results
- Drives rapid success with managed services, getting Arpeely on track to earn millions of $ per month.
- Opens new business for S&P 50 companies by deploying bidding strategies based on terabytes of BigQuery data
- Keeps team lean despite global reach by testing large matrices of machine learning models on live traffic using GKE and Cloud ML Engine
- Empowering engineers with a lean global media operation consisting of an R&D team of less than 25 people
Powering millions of real-time ad predictions per second
Advertising is one of the world’s most sophisticated industries. Behind the catchy slogans and sleek images is an army of data scientists crunching media-buying algorithms that get those ads seen by the people they’re designed for. Today, the financial market that has developed to help advertisers secure the best slots is almost endless. Real-time bidding (RTB) allows media buyers to bid on an ad impression. If they win the bid, their ad is instantly displayed in that slot.
Israeli ad-tech startup Arpeely has supercharged this media-buying process. It cherry-picks traffic using novel machine learning media acquisition algorithms and unorthodox feature engineering techniques.
“With our re-imagined RTB stack, we're able to create tailored strategies and machine learning pipelines for each partner and client,” explains Michael Vainshtein, CTO and co-founder of Arpeely. “Our algorithms, like water, can trickle into hidden market opportunities skipped by the rest of the industry that uses far less granular tools. For example, our engine can automatically zero-down to target only New York early adopters, reading positive-sentiment Tesla articles when it's raining in Manhattan. That’s how specific we can be.”
“We chose Google Cloud because our goal is to keep our team small and to scale without limits. We remain an R&D team of less than 25, but we’re responsible for huge global ad operations running 24/7. The scalability and agility of Google Cloud enables that.”
—Michael Vainshtein, CTO and co-founder, ArpeelyGoogle Cloud is intrinsic to Arpeely’s makeup
Arpeely has come a long way since it was bootstrapped and launched on Google Cloud in 2017. Connected to the world’s largest advertising exchanges, it achieved multimillion dollar revenues in its first year of trading. By now, every single user in the U.S. has passed through Arpeely’s servers at some point, and it processes 20 billion ad impressions a day, while delivering millisecond predictions per ad view.
“We are a cloud-first company that started off with Google Cloud,” says Vainshtein. “We chose Google Cloud because our goal is to keep our team small and to scale without limits. We remain an R&D team of less than 25, but we’re responsible for huge global ad operations running 24/7. The scalability and agility of Google Cloud enables that.”
The fact that Google Cloud offers everything needed to build and scale, leveraging the cloud, also appealed to the cash-strapped startup in its early days. It didn’t want to spend time and money creating products and services that were available off the shelf from Google. “I have always believed that we should build on top of what is already tried and tested and available. Why develop complex services ourselves, when they’re ready for us to use? That’s our approach to Google Cloud,” says Daniel Sirota, VP R&D and co-founder of Arpeely. “Not only does Google Cloud give us the tools we need, it also gives us support when required and has become an ally as we grow the business.”
“We are now serving millions of ad predictions per second across our multi-region GKE load balancers, and our database pipelines can query and run machine learning models on terabytes of data, thanks to BigQuery. GKE and BigQuery allow us to focus on developing the business, not the tools behind it.”
—Michael Vainshtein, CTO and co-founder, ArpeelyGoogle Cloud provided the road map that drove Arpeely’s rapid success. Like many early-stage startups, Arpeely launched with App Engine, which allowed it to quickly iterate in the early days. As volumes scaled up quickly, Arpeely then moved on to the fully managed Google Kubernetes Engine (GKE) and BigQuery.
BigQuery is at the heart of Arpeely’s data warehouse. All its analytics, business metrics, and third-party integrations are streamed to BigQuery, where the data is aggregated. Insight tables are then created in a unique way that does not require ETLs and strategies and models are deployed seamlessly to the Memorystore. This allows analysts and data scientists to easily deploy features and quickly iterate strategies on streams of live traffic.
“We are now serving millions of ad predictions per second across our multi-region GKE load balancers, and our database pipelines can query and run machine learning models on terabytes of data, thanks to BigQuery. GKE and BigQuery allow us to focus on developing the business, not the tools behind it,” says Sirota.
Automatic scaling keeps the workforce lean
Arpeely began with a handful of single-click clusters on GKE. As demand grew, the system automatically scaled up to deploy several hundred nodes. ”Google Kubernetes Engine is our bidding and data engine, we use it extensively. From a DevOps point of view, it means that just one or two people can create self-scaling machines that work day and night. It’s amazing,” says Sirota.
BigQuery, combined with the automated relational database provisioning service, Cloud SQL, helps Arpeely to reign in its teams and stay agile too. Meanwhile, AutoML scales the training of the machine learning models that Arpeely creates as it continually looks to fine-tune its bidding systems. “By automating our machine learning training models, we can test multiple models at the same time, which means that our data scientists don’t have to run each one separately,” says Sirota.
All this lean infrastructure built on Google Cloud explains why when external developers visit Arpeely’s new Tel Aviv-based headquarters, they are often surprised at how few engineers are needed to keep this global data machine going.
“By leveraging pre-built components, Google Cloud lets us stand on the shoulders of giants, so our small team can turn big levers that make great things happen. The successful business we have today would not have been possible without its easily deployable scalable tools and infrastructure.”
—Daniel Sirota, VP R&D and co-founder, ArpeelyThe hidden engine
“We have a lot of Google Cloud tools running quietly in the background that really allow us to concentrate on our unique value proposition,” says Sirota. These include Pub/Sub, which provides messaging and ingestion for Arpeely’s event-driven systems and streaming analytics; Memorystore, which reduces latency and builds application caches; and Cloud Storage, which offers advanced storing capabilities.
“By leveraging pre-built components, Google Cloud lets us stand on the shoulders of giants, so our small team can turn big levers that make great things happen. The successful business we have today would not have been possible without its easily deployable scalable tools and infrastructure,” says Sirota.
With the machinations taken care of, Arpeely is now concentrating on growing the business even further. Vainshtein concludes: “We’re now turning over many millions of dollars a month, and our next goal is to grow by a factor of 10. With Google Cloud as our infrastructure, we are confident that we can do that without having to grow our team 10 times, too.”
Tell us your challenge. We're here to help.
Contact usAbout Arpeely
Data science startup Arpeely uses machine learning and feature engineering to discover hidden opportunities in online advertising. It processes dozens of billions of predictions daily and cherry-picks traffic based on in-app or post-conversion behavior KPIs.
About DoiT International
DoiT International is a global tech consultancy that helps businesses build and scale cloud solutions.