Brave Software improves its web browser using machine learning on AWS

Through their passions for user privacy and browser performance, two tech pioneers have learned how to harmoniously unite advertisers and users while supporting content contributors online.

Brave logo Industry SaaS & ISV, Technology Challenge

Brave built a machine-learning model to classify websites for browser advertising and required a robust way to deploy, manage and audit its machine learning models for its 30 million monthly active users and 1 million+ verified publishers.

Solutions Artificial Intelligence & Machine Learning Platforms Amazon Web Services (AWS)

The challenge of scaling an innovative new browser

With several popular web browsers entrenched in the marketplace, it might seem like the world doesn’t need another one. But the Brave browser was born out of a perceived need to better align user privacy with content creator contributions and advertiser investments. Brave gives users a fast and private web experience, helps advertisers achieve better conversions and helps publishers increase their revenue share.

Brave’s machine learning functionality was designed to make predictions and classify websites into different categories. Based on the content of the sites, Brave can automatically determine what type of websites they are and predict how they will behave from an advertising perspective.

The browser has a unique user attention model for realizing revenue from advertising. Brave combines privacy with a Basic Attention Token (BAT) blockchain-based digital advertising platform. Users can opt-in to privacy-respecting ads that reward them with BAT, an ERC20 token built on top of Ethereum, which they can redeem to tip or contribute to publishers and other content creators.

In a short timeframe, the browser’s unique data privacy features, revenue model and rewards system attracted over thirty million monthly active users and nearly one million verified publishers. In mid-2020, Brave focused on fulfilling its promise of supporting advertisers and content creators and to sustain its early growth trajectory. While the web browsing technology was solid, its machine-learning foundation was not yet strong enough to scale efficiently.
 

“We are incredibly satisfied with Onica and the AWS Jumpstart project. Thanks so much to AWS for offering this to us. We’re appreciative right up to the CEO. It substantially improved the way we created and deployed new models, which has helped us to be much more responsive to advertisers’ needs.”
Jimmy Secretan, VP of Services and Operations, Brave Software
Brave

Improving machine learning predictions

The Brave team of two data scientists and a DevOps specialist created the browser functionality with a fairly basic deployment model. But in order to improve the local machine learning models advertisers depended on, the model pipeline had to be much more robust and scalable.

The first step for Brave Software was to move its machine learning model creation to the cloud. Brave worked with Amazon Web Services (AWS) to make the move. When it came time to bring more power to its machine learning models, the AWS representative suggested working with a specialist in developing machine learning architectural patterns. The rep recommended Onica (a Rackspace Technology company). Onica is a leading AWS Partner Network (APN) Premier Consulting Partner.

“I recommended Onica as a partner, based on their established expertise on AWS, and especially their past successful MLOps Foundations deployments for machine learning operations,” said Brennan Demro, an account manager at AWS. “What’s more, Onica takes a unique approach to working backward from the customer’s goals and building an engagement model that fits the customer to ensure success. This approach is exactly what Brave Software needed.”

Onica is one of only a few service providers to have achieved AWS Machine Learning Competency status. This designation recognizes Onica for demonstrating deep expertise in machine learning experience on AWS, and for delivering its solutions seamlessly in the cloud, solving organizations’ data challenges, and enabling machine learning and data science workflows. The unique combination of expertise in AWS services and machine learning made Onica an ideal partner for Brave.

Brave

"I recommended Onica as a partner, based on their established expertise on AWS, and especially their past successful MLOps Foundations deployments for machine learning operations.”

Brennan Demro, Account Manager, AWS
Brooks Macdonald

Leveraging AWS Services and MLOps Foundations

Before working with Onica, Brave’s process involved several manual steps, including evolving the machine learning model on a local computer, uploading the model to AWS S3 and pushing it out onto a CI/CD pipeline. Also, training the models took several days. Brave needed a more robust pipeline and fully automated processes.

First, Onica modified the cloud infrastructure to prepare for automating the machine learning functionality using a wide range of AWS services, including AWS SageMaker for executing machine learning model training.
AWS Lambda invoked Jenkins endpoints. AWS CodePipeline provided a CI/CD service to continuously train and deploy the machine learning models. AWS CodeCommit provided a source code repository. AWS Systems Manager provided storage at the perimeters, and Amazon S3 stored machine learning training data.

The project was funded through AWS Jumpstart, which helps start-up companies launch new services like SageMaker. Onica is one of the few service providers qualified to provide services for Jumpstart partnerships.

To create a structured way of deploying machine learning models, Onica employed its own MLOps Foundations, which are cloud-based pipelines that simplify machine learning operations by ensuring consistent and traceable model development, training and deployment. It provides an architecture pattern for companies that need to bring that CI/CD pipeline rigor and best practices to their machine learning deployments.

MLOps empowers data scientists to be more efficient and productive. Automating machine learning model creation also supports the privacy component that Brave brings to the table, because power models allow data to stay on the user’s device, without sending user data to the cloud.

“We are incredibly satisfied with Onica and the AWS Jumpstart project,” said Jimmy Secretan, Brave’s VP of Services and Operations. “And we appreciate AWS for offering it to us — right up to our CEO. It substantially improved the way we created and deployed new models, which has helped us to be much more responsive to advertisers’ needs.

Completing the project ahead of schedule

While the infrastructure component of thedevelopment project was straightforward, Brave had an aggressive deadline of six weeks. Plus, Brave modified the deployment model midstream and Onica had to shift gears.

Another challenge was modifying the existing machine learning model, which was a natural language processing model. It needed to be optimized to run on the cloud in a reasonable amount of time. This adjustment helped keep costs manageable.

The Onica team also trained the Brave team on how to use the unique MLOps framework, including how to deploy and troubleshoot it. Even with the needed changes to the model’s architecture, Onica completed the deployment ahead of schedule.

Making the difference in growth potential

Along with critical improvements in its machine learning models, Brave also gained several benefits from its Onica partnership. The model design through MLOps Foundations helped Brave accelerate training, publishing and deploying new models to provide better predictions for browser advertisements.

Model training was reduced from several days to six hours. Also, the streamlined deployment process requires only a few hours instead of a few days, as it previously had. And infrastructure costs dropped by 50%, thanks to purchasing spot instances on an AWS auction service instead of purchasing a dedicated platform.

Brave’s software quality has improved thanks to a new notification and approval process that it didn’t have before. Now, when machine learning model training is complete, the team can review the metrics and accept or reject the updates before they deployed to production.

“Having this essential QA functionality helps us deliver more robust browser software to our users.” said Secretan.

Model training was reduced from several days to six hours, deployments are faster and infrastructure costs were reduced by 50%.

About Rackspace Technology

Rackspace Technology is the multicloud solutions expert. We combine our expertise with the world’s leading technologies — across applications, data and security — to deliver end-to-end solutions. We have a proven record of advising customers based on their business challenges, designing solutions that scale, building and managing those solutions, and optimising returns into the future.

As a global, multicloud technology services pioneer, we deliver innovative cloud capabilities to help customers build new revenue streams, increase efficiency and create incredible experiences. Recognised as a best place to work, year after year, by Fortune, Forbes, Great Places to Work and Glassdoor, we attract and develop world-class talent to deliver the best expertise to our customers. Everything we do is underpinned by an obsession with our customers’ success — our Fanatical Experience™ — so they can work faster, smarter and stay ahead of what’s next.

Let’s Talk Strategy

Tell us a little about your challenges and we’ll contact you.