PGA TOUR calls every shot in real time with Amazon Nova 2 Lite

Professional golf tournaments present a complex coverage challenge. Each tournament spans four days across hundreds of acres of golf course terrain, with more than 150 players competing simultaneously and hitting over 30,000 shots, most of which are not covered by broadcast television.

With this scale, the existing technology didn’t consistently deliver a complete fan experience. While tournaments generate large volumes of real-time data, fans following players outside the top of the leaderboard often received only raw metrics such as yardage, distance, and result, without context needed to understand the shot’s significance within the round.

At the same time, commentary was unavailable for many of the PGA TOUR’s 40-plus simultaneous “Every Shot Live” streams, as assigning dedicated commentators to each feed was not practical.

Delivering meaningful, real-time commentary across every player required more than scale. It required the ability to interpret performance correctly and apply the right context to each shot—something traditional workflows and out-of-the-box language models struggle to achieve. The PGA TOUR sought a way to generate accurate, context-specific commentary that reflects both the current shot and its impact on what comes next, creating a more engaging experience across the full field. 

Trey Pezzetti, AI product manager, PGA TOUR, says, “We wanted to tell the story behind every shot that was hit and help fans understand what happened and why it mattered.”

The PGA TOUR developed an automated AI shot commentary system on AWS to generate real-time, context-aware commentary for every shot across its tournaments. The system was initially built using Anthropic Claude Haiku and later optimized with Amazon Nova 2 Lite on Amazon Bedrock for performance and cost efficiency at scale.

On-course tracking systems, including Trackman radar units and camera feeds, capture structured shot data such as player identity, distance, ball trajectory, and landing position. This data is processed through PGA TOUR’s scoring systems and streamed into a real-time event pipeline.

A rules-based context layer then selects and enriches each shot’s data with relevant context, including leaderboard position, player statistics, and historical performance stored in Amazon DynamoDB. This layer determines the most relevant data points for each shot and applies contextual logic to ensure performance is interpreted correctly. Curated prompts and “gold standard” examples are used to guide the model in generating commentary that reflects both the current shot and its significance within the round.

To operationalize this workflow, shot events flow through Amazon Simple Notification Service (Amazon SNS) and Amazon Simple Queue Service (Amazon SQS) into AWS Lambda, which invokes Amazon Bedrock for generation, integrating seamlessly into the PGA TOUR’s existing system architecture.

 

To maintain accuracy at scale, the PGA TOUR implemented a validation layer using a separate LLM-as-a-judge approach. After commentary is generated, the system extracts key factual elements such as player name, yardage, and shot data, and compares them against the original structured shot data used in the prompt. Commentary is published only when these core details match the source data, ensuring factual consistency before being distributed to front-end systems, including TOURCast, the PGA TOUR’s Emmy Award winning 3D interactive live shot-tracking experience.

Accuracy is sustained over time by treating AI prompts as version-controlled assets, stored in repositories and tested across local, staging, and production environments. Each update is evaluated using validation checks and qualitative assessments for clarity and consistency. “Amazon Nova 2 Lite integrated quickly into our existing system and met our performance requirements immediately, making it the right fit for generating shot-level commentary,” says Pezzetti.

The PGA TOUR now delivers real-time, validated commentary for every shot across each tournament, expanding coverage from a limited broadcast subset to more than 30,000 shots per event. Generated in less than 10 seconds after each shot, the system provides contextual commentary across all 150-plus players, extending coverage previously reserved for featured broadcast groups.

This scale is supported by a lower-cost model configuration. Moving from Anthropic Claude Haiku to Amazon Nova 2 Lite reduced model inference costs by 40 percent helping the PGA TOUR support full-field coverage by automating contextual analysis, commentary generation, and validation workflows.

Beyond coverage and cost improvements, the platform now serves as a foundational layer for the PGA TOUR’s broader AI initiatives. The system processes more than 91 million statistics and 300 data points per shot, combining rules-based contextual analysis with AI-generated commentary to support real-time fan experiences across PGA TOUR digital platforms. Shot-level commentary also powers headline and description generation for video highlights. Reusable prompts, validation workflows, and deployment environments further enable the PGA TOUR to extend the solution more efficiently across additional organizations and use cases.

Pezzetti says, “With Amazon Nova 2 Lite on Amazon Bedrock, we can cover every shot, every player, every hole, every week at a cost that makes it viable to run continuously and build on top of.”