Streamline and Standardize the Complete ML Lifecycle Using Amazon SageMaker with Thomson Reuter

Solution | Scaling the Enterprise AI Platform across TR Using Amazon SageMaker

To create a customized Enterprise AI Platform, TR needed to accommodate a variety of AI use cases, solutions, and AI practitioners’ personas. It also needed to consider scalability, flexibility, governance, and security throughout the ML lifecycle, from model training and deployment to monitoring and explainability.

For experimentation and training, TR needed secure access to data in the cloud to accelerate the development of AI solutions. Using the Enterprise AI Platform, it can quickly spin up ML workspaces based on AWS CloudFormation infrastructure, which speeds up cloud provisioning with infrastructure as code. These workspaces can handle heavy computational workloads and provide access to tools such as Amazon SageMaker Notebooks, which offer fully managed notebooks for exploring data and building ML models. By incorporating purpose-built ML tools with data scientists’ workflows, TR can efficiently run experiments, work on advanced ML projects, and deal with large volumes of data. For example, it analyzed over two million audio files to identify common customer complaints and helped an 11-person team securely and efficiently collaborate on a document-analysis project. “We’ve now streamlined the process for how we create and set up ML resources,” says Dave Hendricksen, senior architect at TR. “In the past, creating an account would take 2–3 months. Now, we can provision one in 2 or 3 days.”

When ML models are ready for deployment, TR uses multiple services based on whether a model is deployed in TR’s products or is destined for internal use. “To deploy models that are going into our products, our product engineering team often uses Amazon SageMaker endpoints,” says Apazoglou. “For teams that are creating AI for internal consumption, we have developed a deployment service that codes Amazon SageMaker bots to run inferences for the models on a periodic schedule.” To monitor its ML models for drift or potential bias and to provide explainability of generated insights, TR uses Amazon SageMaker Model Monitor, a service that keeps ML models accurate over time. It also relies on Amazon SageMaker Clarify, which detects bias in ML data and explains model predictions. By extending these solutions, the company can schedule and evaluate AI models’ performance according to predefined metrics and receive notifications whenever bias or drifts are detected.

Finally, the Enterprise AI Platform’s Model Registry provides a central repository for all TR AI/ML models. This component is partly based on Amazon SageMaker Model Registry, which companies use to catalog models for production, manage model versions, and associate metadata—such as training metrics—with a model. Using this service, the company makes ML models that are developed across multiple AWS accounts and are owned by different business units available to view and potentially to reuse, making it simple for teams to collaborate. TR also gains transparency and orchestration of model workflows as well as a centralized view of models’ metadata and health metrics.

On AWS, TR can better meet its ML model governance standards and empower its data scientists to build innovative, secure, and powerful AI services to serve end users. The company is using the solution at scale across its entire enterprise and has seen widespread adoption across its data science teams. In fact, more than 150 AI professionals are using the solution.