PodFlower is a sophisticated multi-agent system built with Google's Agent Development Kit (ADK) that automates the complete end-to-end workflow for Japanese podcast production. From raw audio processing to social media distribution, this system demonstrates the power of AI agent orchestration in solving real-world automation challenges.

This project showcases:

• ✅ Multi-agent orchestration using ADK Workflow Agents
• ✅ Complex process automation with 10 specialized agents
• ✅ Real-world impact for Japanese podcast production
• ✅ Zero-touch workflow from recording to distribution

python pipelines/full_workflow.py sample_episode/

This single command processes a complete podcast episode with zero manual intervention.

PodFlower uses 10 specialized agents orchestrated through ADK's workflow patterns:

graph TB
    A[🎤 Recorder Agent] --> B[✂️ Filler Removal Agent]
    B --> C[🔗 Concat Audio Agent]
    B --> D[📝 Title Generator Agent]
    B --> E[📢 Ad Break Detector]
    C --> F[🎛️ Mastering Agent]
    F --> G[📦 Export Package Agent]
    D --> G
    E --> G
    G --> H[🚀 Vercel Deploy Agent]
    G --> I[📰 WordPress Publisher]
    G --> J[🐦 X Post Agent]

• Python 3.11+
• FFmpeg
• Google Cloud account
• Google API key (from Google AI Studio)

# Clone the repository
git clone <repository-url>
cd podflower

# Install dependencies
make install

# Configure environment
cp env.example .env
# Edit .env with your API keys

# Required
GOOGLE_API_KEY=your_google_api_key
GOOGLE_CLOUD_PROJECT=your-project-id

# Optional (for full functionality)
VERCEL_TOKEN=your_vercel_token
WORDPRESS_USERNAME=your_username
WORDPRESS_APP_PASSWORD=your_password
X_CONSUMER_KEY=your_x_api_key
# ... see env.example for complete list

podflower/
├── agents/                  # All ADK agents
│   ├── recorder/           # Audio file detection
│   ├── filler_removal/     # Japanese filler word removal
│   ├── concat_audio/       # Intro/outro concatenation
│   ├── title_notes/        # LLM-powered content generation
│   ├── ad_break/           # Topic-shift detection
│   ├── mastering/          # Audio normalization
│   ├── export_package/     # Final package creation
│   ├── deploy_vercel/      # Static site deployment
│   ├── wordpress_publish/  # Content management
│   └── post_to_x/          # Social media posting
├── pipelines/
│   └── full_workflow.py    # Main orchestration pipeline
├── tests/                  # Comprehensive test suite
├── sample_episode/         # Sample input directory
├── assets/                 # Intro/outro audio files
└── SPEC.md                # Complete technical specification

1. Recorder Agent - Detects new Zoom/Riverside recordings
2. Filler Removal Agent - Uses Google Cloud Speech-to-Text for Japanese filler word removal
3. Concat Audio Agent - Adds intro/outro using FFmpeg
4. Mastering Agent - Normalizes to -16 LUFS / -1 dB peak

• Title Notes Agent - Generates 5 Japanese titles + Markdown show notes using Gemini
• Ad Break Agent - Detects topic shifts using sentence transformers

• Export Package Agent - Creates timestamped episode directory with metadata

• Vercel Deploy Agent - Triggers static site redeploy
• WordPress Publisher - Posts with Unsplash featured image
• X Post Agent - Announces episode with proper Japanese formatting

# Run all tests
make test

# Run specific test types
make test-unit      # Unit tests
make test-e2e       # End-to-end pipeline tests

# Performance benchmarks
make bench

This system is specifically designed for Japanese podcast content:

• Filler Word Detection: Removes "えーと", "あのー", "まあ", etc.
• Japanese Speech-to-Text: Uses Google Cloud Speech-to-Text with ja-JP model
• Native Title Generation: Creates natural Japanese episode titles
• Cultural Ad Placement: Respects Japanese podcast conventions

# Format code
make format

# Type checking
make type-check

# Development server
make demo          # Run with verbose logging

• Coverage: ≥ 90% test coverage
• STT Accuracy: ≤ 5% word error rate after filler removal
• Type Safety: mypy --strict compliance
• Performance: Complete workflow in < 10 minutes for 60-minute episode

python pipelines/full_workflow.py sample_episode/

make docker-build
make docker-run

See deployment documentation for cloud deployment options.

• Multi-agent orchestration using all three ADK agent types
• Language-specific AI processing for Japanese content
• Real-time audio DSP with professional mastering
• Clean, typed, well-documented code following SPEC

• Solves real workflow pain for Japanese podcast creators
• Zero-touch automation from recording to social media
• Cultural adaptation for Japanese language nuances
• Professional audio quality matching industry standards

• One-command demo: python pipelines/full_workflow.py sample_episode/
• Comprehensive documentation with clear setup instructions
• Complete test coverage including end-to-end pipeline tests
• Architecture diagrams showing agent orchestration

• ✅ ADK contribution potential - demonstrates advanced agent patterns
• ✅ Google Cloud integration - Speech-to-Text, Storage, potential Vertex AI
• ✅ Production-ready - error handling, logging, monitoring

We welcome contributions to PodFlower! This project demonstrates advanced ADK patterns and could serve as a reference implementation for:

• Multi-agent workflow orchestration
• Language-specific AI processing pipelines
• Real-world automation use cases
• Japanese language AI applications

This project is intended as a demonstration of ADK capabilities for automating complex, real-world workflows.

• ADK Documentation: https://google.github.io/adk-docs/
• Demo Video: [Coming soon]
• Architecture Blog: [Coming soon]

Built with ❤️ to demonstrate the power of AI agent collaboration in solving real-world automation challenges.