Articul8 AI Improves Productivity by 35% and Helps Customers Reduce AI Deployment Time by 4x Using Amazon SageMaker HyperPod

Launched in January 2024, Articul8 AI offers a full-stack, vertically optimized generative AI solution, purpose-built to meet the four core needs of its customers: scalability, speed of deployment, security, and sustainable costs. The company offers generative AI technology that includes capabilities such as multi-modal understanding, dynamic reasoning and decisioning, autonomous actioning, explainability, auditability, and domain-specific/enterprise-specific model development.

Articul8 AI’s unique, self-contained software model helps customers rapidly build, deploy, and manage enterprise-grade generative AI applications while keeping their data within their own virtual private clouds.

To build its generative AI platform, the Articul8 AI team needed to quickly and efficiently deploy large scale infrastructure resources to manage workloads for internal research and development as well as production deployment.. Articul8 AI also faced the challenge of managing GPU clusters without a dedicated infrastructure team. The company needed a solution that would increase compute cluster resiliency along with high availability and automated management while also standardizing its technology stack so that it could deploy seamlessly in its customers’ infrastructure. Articul8 AI chose to build its solution on AWS because of its reliability, scalability and the proactive support provided by the AWS service teams.

One of the most computational- and time-intensive jobs for Articul8 AI was pretraining and fine-tuning the various large language models including domain-specific/enterprise-specific models that form the backbone of Articul8 AI’s generative AI platform. At Articul8 AI’s scale, any kind of performance degradation was unacceptable because it could negatively impact the cost and time it took to train the models. So, the company sought a high-performance solution that could accelerate its model training time. The company experimented with several solutions but ultimately adopted Amazon SageMaker HyperPod, which typically reduces the time to train foundation models by up to 40 percent with a purpose-built infrastructure for distributed training at scale.

Articul8 AI worked alongside AWS to collaboratively problem-solve and go from concept to deployment, which helped the company quickly move forward with their training workloads. “When we first started, we wanted to see how this solution would grow and adapt to our needs, and we couldn’t be happier with the support we received from AWS,” says Felipe Viana, head of applied research at Articul8 AI. “Getting the support we needed in a timely manner helped us gain confidence in the Amazon SageMaker HyperPod service and sped up our AI research work.”

Articul8 AI spent over 18 months building its autonomous, vertically integrated generative AI platform in stealth mode before publicly launching in January 2024 with two main product offerings with full enterprise support. The first is the Express Bundle, which is a preconfigured product for specific use cases, such as knowledge discovery, that do not require training or fine-tuning capabilities for customers. The second option is the customizable Premium Bundle that enables continued pretraining or fine-tuning of domain-specific/enterprise-specific models along with a host of other features and AI packs required for production-grade enterprise deployments. The solution stack for both these options is built on AWS and uses Amazon SageMaker HyperPod for model training and fine-tuning.

Using Amazon SageMaker HyperPod made it possible for Articul8 AI to enhance performance and get its solution to market quickly. Using Amazon SageMaker HyperPod removed the need for the company to manually manage its high performance compute clusters. As a result, Articul8 AI got its cluster usage up to nearly 100 percent. In addition, the company saw its productivity improve by up to 35 percent. Amazon SageMaker HyperPod also monitors cluster health and automatically replaces faulty nodes. This reduced Articul8 AI’s time to train and made the deployment frictionless for researchers, improving overall productivity. “The time we’ll save by not managing the clusters ourselves will really add up,” says Arun Subramaniyan, founder and CEO at Articul8 AI. “We’re investing in technologies such as Amazon SageMaker HyperPod up front to help our teams operate efficiently and to scale up our operations rapidly.”

Articul8 AI also integrated Amazon SageMaker HyperPod with other AWS services, including Amazon Managed Grafana, which offers scalable and secure data visualization for operational metrics, logs, and traces. The company further used Amazon SageMaker HyperPod lifecycle scripts to customize their cluster environment and install required libraries and packages including open-source exporters. With these additions, Articul8 AI achieved near real-time observability of its GPU resources in a single-pane-of-glass dashboard.

The prebuilt, out-of-box solution helps Articul8 AI’s customers accelerate their generative AI journeys with access to a collection of AI models as well as a multitude of other features, all deployed and operated autonomously within the customer’s secure environment. Articul8 AI’s customers experience up to four times faster deployment times compared with other generative AI solutions, cutting a potentially months-long process down to mere weeks for certain use cases. Articul8 AI’s optimizations on AWS and differentiated business model also help customers to reduce their total cost of ownership by up to five times when deploying their generative AI applications in production at scale.