Summary

Introduction

• The NFL is using generative AI (Gen AI) to streamline research, content creation, and discovery.
• The goal is to enable content creators to find footage faster by translating natural language queries into database-understandable requests.

Playbook Pro Demo

• Playbook Pro is a chatbot-based solution that allows users to search for plays based on various criteria, such as player performance, completion probability, and more.
• The solution uses a websocket API to provide real-time updates on the search process and allows users to preview and access the relevant video clips directly.
• The system also supports conversational interactions, where users can refine their queries and get additional information.

Motivation and Approach

• The NFL team faced challenges with their existing tools, which required extensive manual effort and familiarity with complex filtering mechanisms.
• The AWS Gen Innovation Center and the NFL team took a systematic approach to identify the right problem to solve and the technical feasibility of using Gen AI.
• The key steps included:
  • Aligning with stakeholders to understand the business problem
  • Diving deeper into the use case, data availability, and user experience
  • Building a proof-of-concept (POC) to validate the approach and the model's performance
  • Iterating quickly and incorporating user feedback to refine the solution

Proof-of-Concept Development

• The main challenges were:
  • Translating natural language queries into database-understandable requests
  • Accurately identifying relevant entities (players, teams, etc.) and associating them with the correct data
  • Mapping actions and constraints to the appropriate API filters
• The team leveraged techniques such as:
  • Custom entity recognition using traditional NLP and large language models
  • Step-by-step reasoning in the language model to improve accuracy and transparency
  • Incorporating both correct and incorrect examples to train the model

Moving to Production

• The AWS Professional Services team took the POC and developed a production-ready solution using an "agentic workflow" approach.
  • This involves a graph-based architecture with nodes (actions), edges (relationships), and state (information passed between steps).
• Key aspects of the production system include:
  • Assessing user intent upfront to enable conversation and deflect irrelevant requests
  • Leveraging a semantic vector database to retrieve and reuse similar past executions
  • Implementing self-correction loops to handle API call failures
  • Summarizing results and providing links to the media asset management system

Impact and Future Work

• The Playbook Pro solution has significantly reduced the time it takes for content creators to find relevant footage, from 10 minutes to 30 seconds.
• This allows them to spend more time on the creative aspects of their work.
• Future plans include:
  • Expanding the coverage of the Next Gen Stats API
  • Incorporating biographical and historical player data
  • Developing a fan-facing application