Transforming PayPal’s AI Culture to Enable Intelligent Business Decisions

Image by Lagos Techie from Unsplash

At PayPal, we believe technology is most powerful when it empowers everyone. This is especially true when it comes to broadening Artificial Intelligence (AI)/Machine Learning (ML) innovation across the company. In an effort to democratize AI, an end-to-end data exploration, feature engineering, model training, and inferencing platform that is available to all business domains and cross-functional users are critical to PayPal’s success.

Democratize AI to enable broader business transformation

In the past few years, High-Performance Computing (HPC) hardware is increasingly powerful, big data is easily accessible, and advanced algorithms or autoML tools are available through open-source packages, vendor tools, and public cloud services. Due to the wide range of business demands and the shortage of highly skilled data scientists, the need to democratize AI, lower the entry of AI and make AI available to the larger user community in PayPal is more prominent.

For example, finance business analysts create models to adjust pricing strategies based on customer sentiment to increase revenue. Developers in customer success build models to predict potential customer complaints, then route the incoming customer calls to the best serving channel to increase customer satisfaction. Site reliability engineers train models to predict potential capacity or network issues and proactively correct problems before problems happen to avoid system loss.

Figure 1: AI/ML Use Cases in PayPal’s Business Domains

Although not all use cases could generate hefty dollar value, the collective value of use cases from all domains is significant to the company. To truly transform our business, our AI culture needs to change so that everyone can innovate, and every business decision is AI-driven.

What is the AI/ML Platform?

PayPal’s AI/ML platform allows users to develop, deploy, and maintain machine learning and deep learning models over time. It’s a set of managed services that work holistically and cohesively to enable a reliable, scalable, secure, and self-service experience.

The platform is built by customizing and integrating homegrown, open-source, and vendor tools and cloud services in a hybrid cloud environment. This ensures the platform can meet PayPal’s business needs and InfoSec requirements.

Figure 2: AI/ML Platform

AI/ML Platform also aims to help PayPal’s AI community to discover and share ML assets like data, features, and models, etc. This would enable the community to innovate together to further reduce the AI time-to-market in all business domains.

The self-service console simplifies the ML workflows based on the users’ goals and/or ML proficiency. For example, with a few clicks, a citizen data scientist uses the autoML capability of the platform to select the best-performed model from many other auto-trained models. The platform then guides her to deploy, serve, and monitor the model in production. Here, feature selection and engineering, model training, experiment and evaluation, and hyperparameter tuning are all done automatically in the platform.

There are seven user personas that the AI/ML Platform creates value for:

Figure 3: User Personas of the AI/ML Platform (avatar images designed by freepick)

The AI/ML Platform also supports expert data scientists. For example, the majority of the mission-critical models were trained in PayPal Notebook which is part of the AI/ML Platform. It manages deep learning workflows and enables distributed training by splitting a training job into many GPU nodes for data or network parallelization. The trained model is then simulation tested after being pushed to a model management repository. After shadow auditing, the candidate model is then deployed to the real-time production environment.

Standardize ML Development Lifecycle to increase AI/ML agility

Although the creation, test, and execution of ML models are very different from software code, the development lifecycle concept, and Continuous Integration/Continuous Deployment (CI/CD) rigor are no different or even more stringent due to the reproducible needs and the repetitive and experimentation nature of ML.

The platform hides the complexities by providing a self-service portal that takes a data scientist or developer to release his/her model as easy as a software engineer to release code. It aims to simplify and streamline the end-to-end ML workflow from data analysis, model development, to productization.

Figure 4: PayPal’s ML Development Lifecycle

Platform Benefits

The benefits of this platform include:

• Increased productivity of all platform users
• Reduced time-to-market of AI/ML solutions
• Intelligent business decisions empowered by ML models
• Continuous value delivery and innovation at scale
• Lowered Total Cost of Ownership (TCO)

By democratizing and integrating AI/ML into workflows and product development, the AI/ML Platform helps to reduce barriers to critical data and algorithms that many domains need to accomplish their business goals. This, in turn, leads to innovation, important insights, and significant gains for PayPal.

Appreciations

Many thanks to the AI/ML Platform team for making this platform from concept to reality. The platform’s creation and evolvement cannot be done without the team’s tireless effort.

Thanks to Linda Merry for editing and making this story publication-ready.