PodFlower: AI-Powered Podcast Automation

PodFlower: AI-powered Japanese podcast automation using 10 ADK agents. Transforms 4-7 hours of manual work into 1 command with professional audio mastering and multi-platform distribution.

Comment

Inspiration

The Japanese podcast market is rapidly growing, but creators face a 4-7 hour manual workflow for each episode: audio editing, content generation, and multi-platform distribution. I saw an opportunity to leverage Google's ADK to build the first comprehensive Japanese podcast automation system, addressing the unique challenges of Japanese speech processing including filler word removal ("えーと", "あの") and cultural content generation.

What it does

PodFlower is a sophisticated multi-agent system that transforms raw Japanese audio into broadcast-ready podcast episodes with a single command. My 10 specialized agents handle the complete workflow: Japanese speech processing with filler word removal, professional audio mastering to -16 LUFS broadcast standards, AI-powered content generation using Gemini, and automated multi-platform distribution to Vercel, WordPress, and social media.

How I built it

I architected PodFlower using all three ADK agent families: LLM agents for Japanese NLP and content generation, Workflow agents (SequentialAgent, ParallelAgent) for complex orchestration, and Custom agents for audio processing and deployment. The system integrates Google Cloud Speech-to-Text for Japanese transcription, Gemini Pro for content generation, FFmpeg for professional audio processing, and deploys live on Vertex AI Agent Engine for production scalability.

Challenges I ran into

The biggest challenge was Japanese language processing - detecting and removing filler words while preserving natural speech flow. I also faced complex state management between 10 agents, ensuring seamless data flow from audio processing through content generation to distribution. Audio mastering to broadcast standards required precise FFmpeg configuration, and deploying the multi-agent system to Agent Engine involved solving serialization issues with complex agent orchestration.

Accomplishments that I'm proud of

I built the most complex ADK implementation in the hackathon with 10 agents using all three agent families. My system achieves 95% time reduction (4-7 hours to 1 command) while maintaining professional broadcast quality. I successfully deployed a production-ready system on Vertex AI Agent Engine and created the first Japanese-specialized podcast automation platform with real business impact.

What I learned

I mastered advanced ADK patterns including complex multi-agent orchestration and state management. I gained deep expertise in Japanese NLP processing and professional audio engineering standards. The project taught me how to design production-ready AI systems that solve real business problems while maintaining technical excellence and user experience.

What's next for PodFlower: AI-Powered Podcast Automation

I plan to expand language support beyond Japanese, add advanced analytics for content optimization, integrate with more podcast platforms, and develop a SaaS offering for podcast creators. I'm also exploring partnerships with Japanese media companies and considering open-sourcing core components to benefit the broader podcast creation community.

Built With

Share this project:

Updates