NHL Live Podcast System

A sophisticated multi-agent architecture that transforms live NHL game data into engaging, real-time commentary using Google's ADK, generates professional audio commentary for NHL games.

Comment

🏒 About the Project

NHL LiveStream Commentary Agent is a real-time AI-powered hockey commentary system that transforms live NHL game data into engaging, professional audio commentary. Built using Google's Agent Development Kit (ADK), this project demonstrates the power of multi-agent AI systems in creating dynamic, contextual sports broadcasting.

💡 What Inspired Us

The inspiration came from watching live sports broadcasts and noticing how commentary can make or break the viewing experience. We wondered: "What if AI could generate professional-quality, real-time commentary that rivals human broadcasters?"

Key inspirations:

  • Real-time Sports Broadcasting: The challenge of creating instant, contextual commentary
  • Multi-Agent AI Systems: Exploring how different AI agents can collaborate seamlessly
  • Google ADK Capabilities: Pushing the boundaries of what's possible with Google's Agent Development Kit
  • Hockey's Fast-Paced Nature: The sport's rapid tempo provides an excellent testbed for real-time AI processing

🎯 What We Learned

Technical Learnings

  • Agent Orchestration: How to coordinate multiple AI agents (Data, Commentary, Audio) in real-time
  • Session Management: Maintaining context continuity across agent interactions to prevent repetitive commentary
  • Data Integrity: Preventing "data leakage" where future game information contaminates early timestamps
  • Progressive Statistics: Building realistic game progression from 0-0 to final scores
  • WebSocket Streaming: Real-time audio delivery for live broadcasting applications

ADK-Specific Insights

  • Sequential Agents: Learned to chain agents effectively using ADK's Sequential Agent pattern
  • Custom Agent Implementation: Built complex orchestration logic using ADK's BaseAgent class
  • Tool Integration: Created sophisticated tools for momentum analysis, player information, and audio processing
  • Session Persistence: Implemented proper session management for production deployments

Sports Data Challenges

  • NHL API Complexity: Understanding the intricacies of real-time sports data
  • Temporal Consistency: Ensuring commentary flows naturally without timeline conflicts
  • Player Name Resolution: Mapping player IDs to human-readable names across different data sources

🔧 How We Built It

Architecture Overview

NHL API → Live Data Collector → Data Agent → Commentary Agent → Audio Agent (Progressive stats) (ADK Analysis) (Two-person) (TTS Stream)

Development Phases

Phase 1: Foundation (Data Agent)

  • Built ADK-powered data analysis agent using google.adk.Agent
  • Implemented tools for momentum analysis and game context extraction
  • Created progressive statistics system to prevent data leakage

Phase 2: Commentary Generation (Commentary Agent)

  • Developed dual-commentator system (Play-by-play + Analyst)
  • Implemented session-aware processing for natural conversation flow
  • Created Alex Chen & Mike Rodriguez personas for authentic broadcasting

Phase 3: Audio Integration (Audio Agent)

  • Built custom ADK agent using BaseAgent class for complex orchestration
  • Integrated Google Cloud Text-to-Speech with dynamic voice styling
  • Implemented WebSocket streaming for real-time audio delivery

Phase 4: Sequential Agent Innovation

  • V2: Clean Sequential Agent wrapper for Data + Commentary
  • V3: Complete pipeline with integrated audio processing
  • Solved agent coordination challenges through custom processing logic

Phase 5: Production Deployment

  • Containerized application with Docker
  • Deployed to Google Cloud Run with auto-scaling
  • Built web demo interface with real-time audio streaming

Key Technologies

  • Google ADK: Agent Development Kit for multi-agent coordination
  • Google Gemini: Large language model for commentary generation
  • Google Cloud TTS: Text-to-speech for audio generation
  • NHL API: Real-time game data source
  • WebSockets: Live audio streaming
  • Python AsyncIO: Concurrent processing
  • Docker & Cloud Run: Containerized deployment

🚧 Challenges We Faced

  1. Agent Coordination Complexity

Challenge: Coordinating three different agents with different response patterns and timing requirements.

Solution: Developed Sequential Agent V3 with custom processing logic that manually orchestrates agent calls rather than relying solely on ADK's built-in Sequential Agent.

# Custom agent orchestration data_response = await self._call_agent_with_message(self.data_agent, message) commentary_input = self._prepare_commentary_input(message, data_response) commentary_response = await self._call_agent_with_message(self.commentary_agent, commentary_input) audio_result = await self.process_commentary_to_audio(commentary_data)

  1. Data Leakage Prevention

Challenge: Early game timestamps were contaminated with final game statistics, making commentary unrealistic.

Solution: Implemented progressive statistics calculation that only uses events up to the current timestamp:

# Filter events by timestamp to prevent future data leakage filtered_events = [event for event in all_events if event.timestamp <= current_time] progressive_stats = calculate_stats_from_events(filtered_events)

  1. Session Management & Memory

Challenge: ADK sessions accumulated too much context, leading to repetitive commentary and memory issues.

Solution: Implemented smart session refresh strategy:

  • Refresh sessions every 8-10 files
  • Maintain recent dialogue history for context
  • Compact narrative state between sessions
  1. Real-time Audio Processing

Challenge: Generating and streaming audio in real-time while maintaining chronological order.

Solution: Built sequential processing pipeline:

  • Process files in strict chronological order
  • Generate audio immediately after commentary
  • Use WebSocket for live streaming delivery
  1. NHL API Data Complexity

Challenge: NHL API returns complex nested data structures with inconsistent player naming.

Solution: Created comprehensive data enhancement system:

  • Player name resolution with fallback strategies
  • Spatial coordinate conversion for context
  • Robust error handling for missing data
  1. Production Deployment

Challenge: Deploying multi-agent system with real-time requirements to cloud infrastructure.

Solution:

  • Containerized with optimized Docker images
  • Used Cloud Run for auto-scaling
  • Implemented proper environment variable management
  • Created comprehensive logging and monitoring

🎉 Results & Impact

Technical Achievements

  • Sub-5 Second Latency: From NHL API to audio output
  • Natural Commentary Flow: Eliminated repetitive AI commentary through session management
  • Production Ready: Deployed and running on Google Cloud Run
  • Scalable Architecture: Handles multiple concurrent games

Innovation Highlights

  • Sequential Agent V3: Novel approach to multi-agent coordination
  • Progressive Statistics: Solved data leakage in sports AI applications
  • Live Game Board: Pure state tracking with anti-hallucination features
  • Real-time Audio Streaming: WebSocket-based live commentary delivery

Demo Capabilities

  • Process any NHL game by ID
  • Generate realistic two-person commentary
  • Stream live audio to web clients
  • Support multiple concurrent games
  • Maintain commentary quality across extended periods

🚀 Future Enhancements

  • Multi-language Support: Expand beyond English commentary
  • Visual Integration: Add player tracking and visual analysis
  • Advanced Analytics: Deeper statistical analysis and predictions
  • Broadcast Integration: Direct integration with streaming platforms
  • Fan Interaction: Real-time fan questions and commentary requests

This project showcases the incredible potential of Google's ADK in creating sophisticated, real-world AI applications that can compete with traditional human-driven processes. The combination of intelligent agents, real-time processing, and cloud deployment creates a foundation for the future of AI-powered sports broadcasting.

Built With

Share this project:

Updates

Submission history