Strava safely transforms the athlete experience using Amazon Bedrock

With tens of millions of activities uploaded daily across more than 50 activity types, Strava’s users generate an enormous amount of data. The platform captures stats such as heart rate and pace, but many users want to understand what these numbers mean for their progress—that is, how their effort compares to their activities in the recent past.

Strava wanted to solve this challenge by providing personalized insights in an authentic, encouraging brand voice to help its users interpret their activities. “To get something that feels unique to each athlete, at our scale and with the diversity of users that we have, can be incredibly challenging from a technical and a quality perspective,” says Katy Evans, director of engineering for AI and ML at Strava.

To explore how to efficiently use generative AI to deliver insights for users at scale, Strava engaged the AWS Generative AI Innovation Center. Together, the teams looked at models from leading AI companies to determine the right large language model for Athlete Intelligence. Ultimately, they chose to work alongside Anthropic, which provided the security, trust, and access crucial to Strava’s development of the new feature. The involvement of Anthropic drove a three-way synergy during the advisory stage and the pre- and postlaunch of Athlete Intelligence. After rigorous in-house testing to determine which model blended cost, latency, and quality of outputs, Strava selected Anthropic’s Claude Haiku in Amazon Bedrock to power the feature.

Users can record activities directly on their phones using the Strava App or by uploading from a connected device. Once they complete their activity, the athlete has immediate access to an activity detail page that presents a summary complete with a map, graphs, and in-depth stats. Behind the scenes, Claude Haiku analyzes these stats in the context of the athlete’s history of activities on Strava and considers the athlete-provided descriptions of what they’ve just done. Within seconds, Athlete Intelligence generates personalized insights in a friendly, engaging brand voice, such as “Breakthrough run with multiple personal bests and setting segment records on a familiar route.”

Using Claude in Amazon Bedrock helps Strava scale seamlessly to 14 million tokens per minute at peak throughput. As a result of this scalability through Amazon Bedrock, Strava can serve more users at once with context-aware insights, accelerating response times.

Because Athlete Intelligence analyzes user-generated content such as activity descriptions and titles, Strava needed robust safeguards to help keep outputs motivating and appropriate for its global community. “Using Amazon Bedrock Guardrails helps us detect and omit harmful information while keeping the tone of responses positive, uplifting, and healthy,” says Alaena Roberds, AI engineer at Strava. “We also had a lot of support from AWS that helped us feel comfortable, which was a good confidence boost.”

After the feature launched, Strava was pleased that more than 80 percent of users who provided feedback indicated that they found the Athlete Intelligence insight to be “very helpful” or “helpful” for them. “If we had self-hosted something like this, that would have taken away valuable time that we could have used to build product features to support customer value,” says Evans. Strava’s developers have access to cutting-edge models and powerful tools and capabilities through Amazon Bedrock so that they can continue to develop even more features for users. “We need to ‘build flexible’ in a fast-moving area,” says Roberds. “Our product is always changing. And we want Athlete Intelligence to be able to change along with it.”

Strava continues to collaborate with the AWS Generative AI Innovation Center and to emphasize reliability as they work to accommodate users with a hunger for increasingly personalized features. “We imagine a world where fitness becomes increasingly personalized, adaptable, responsive, and compassionate,” says Sambrooke. “We want to meet users where they are and help them be athletic and move with the currents of their lives. And when technologies such as Claude by Anthropic in Amazon Bedrock can bring that to the table, that’s very encouraging and affirmative.”