AWS re:Invent 2025 - Powering Prime Video's NASCAR Coverage: ML Fuel Analytics in Action (SPF303)

Powering Prime Video's NASCAR Coverage: ML Fuel Analytics in Action

Overview

• The presentation showcases how the Prime Video innovation team developed a real-time fuel analytics solution to enhance the NASCAR coverage on Prime Video.
• The key challenge was to provide fuel strategy insights to on-air talent and viewers in under 5 seconds, despite the lack of direct fuel data from the NASCAR cars.

Fuel Analytics Approach

• The team took a multi-pronged approach to fuel analytics:
  1. Visual Analysis: Trained computer vision models to extract fuel consumption patterns from video footage.
  2. Physics-based Modeling: Developed parametric models mapping telemetry data to fuel consumption, calibrated using historical race data.
  3. Machine Learning: Built predictive AI models on Amazon SageMaker to estimate fuel consumption from the telemetry data.

AWS Architecture

1. Data Ingestion:

   • Telemetry data from NASCAR's ERDP platform ingested into AWS using Kinesis Data Streams and Fargate tasks.
   • Best practices applied, such as asynchronous processing, adaptive sampling, and monitoring for back-pressure.

2. Real-time Processing:

   • Apache Flink (managed service) used as the core processing engine.
   • Leveraged Flink concepts like key streams, broadcast streams, and tumbling windows to enable low-latency, high-throughput processing.
   • Managed state using Flink's key state primitives and checkpoint/restore capabilities for fault tolerance.

3. Fuel Analytics Integration:

   • The physics-based and ML models were integrated into the Flink pipeline to perform real-time fuel consumption calculations.
   • The processed fuel analytics data was published to an output Kinesis stream.

4. Dashboards and Delivery:

   • AWS AppSync and DynamoDB used to power real-time dashboards for on-air talent.
   • GraphQL subscriptions enabled low-latency, bi-directional updates between the backend and the custom React-based front-end applications.

Key Results and Impact

• The "Burn Bar" feature was launched for the Coca-Cola 600 race, the first time fuel strategy was made visible to broadcasters and fans.
• Achieved 534 million media impressions and reached over 2 million viewers on average.
• Met stringent performance and accuracy KPIs, providing fuel insights to on-air talent in under 5 seconds.

Lessons and Recommendations

• Embrace failure and be open to taking uncharted paths when innovating.
• Focus on scaling successful experiments to unlock real value.
• Empower teams to accelerate experimentation velocity using AWS services and tools.
• Leverage the AWS Well-Architected Framework to guide the design and improvement of solutions.

Technical Details

• Services used: Kinesis Data Streams, Fargate, Amazon Managed Streaming for Apache Flink, Amazon SageMaker, DynamoDB, AWS AppSync, AWS Lambda, Amazon CloudFront.
• Key Flink concepts: Key streams, broadcast streams, tumbling windows, checkpointing, watermarks.
• Kinesis capacity modes: Provisioned, On-Demand Standard, On-Demand Advantage.
• Achieved MVP in 8 weeks, productionized in 12 weeks with a small team.

Business Impact

• Enabled broadcasters to provide deeper, real-time insights on fuel strategy to NASCAR fans, enhancing the viewing experience.
• Unlocked new storytelling opportunities around a critical but previously hidden aspect of NASCAR racing.
• Laid the foundation for further innovation and feature development around NASCAR coverage on Prime Video.

Examples

• Showcased how the "Burn Bar" feature was used by on-air talent to analyze fuel efficiency differences between drivers during the Pocono race.
• Highlighted how the solution provided insights into Chase Briscoe's fuel-saving strategy that helped him win the Pocono race.