PodPilot

An end-to-end automation system that discovers trending podcasts, extracts highlight moments using LLMs, and publishes short-form videos to YouTube.

PodPilot handles everything — from podcast discovery to video editing and publishing — with no manual intervention.

🚀 How It Works

The system operates as a sophisticated pipeline, orchestrated by LangGraph:

• Search Podcasts: Queries YouTube for recent and relevant podcasts, fetching initial metadata.
• Select Best Podcast: An LLM analyzes the retrieved podcasts (excluding previously processed ones) to select the most promising candidate for clip generation. If no suitable podcast is found, it retries with different filters.
• Process Video: Downloads the selected podcast's video and extracts its transcript (captions).
• Fetch Clips (Moment Identification): An LLM processes the transcript in chunks, identifying and detailing engaging "moments" (start time, end time, title, description, keywords).
• Edit Video: For each identified moment, the system:
  • The segment is trimmed with ffmpeg
  • Captions are burned in
  • 9:16 aspect ratio is applied with a blurred background for Shorts-style viewing
• Post Video: Uploads each generated video clip to YouTube with its AI-generated metadata.
• Clean Up & Loop: Clears local data and marks the original podcast as "burnt" (processed) before initiating the next cycle.

Flow

graph TD
    A[Start: Search Podcasts] --> B[Select Best Podcast via LLM]
    B -->|Valid Podcast Found| C[Process Video - Transcript Extraction]
    B -->|No Suitable Podcast| BB[Retry with Different Filter]
    BB --> A

    C --> D[Fetch Clips - LLM Identifies Moments]
    D --> E[Edit Video - Trim & Burn Captions]
    E --> F[Post Clips to YouTube]
    F --> G[Clean Up & Mark Podcast as Processed]
    G --> A

    subgraph Error Handling
        B --> X[Error Handler]
        C --> X
        D --> X
        E --> X
        F --> X
    end

Loading

🛠️ Tech Stack

Component	Purpose
LangGraph	Orchestrates multi-step, stateful workflows
LangChain	Manages LLM calls and prompts
Gemini 2.0 Flash	LLM used for selection and clip extraction
FFmpeg	Trimming, resolution adjustment, and caption embedding
yt-dlp	Podcast discovery and video download
YouTube Data API v3	Video uploading
LangSmith	Observability and debugging of the LLM pipeline

⬇️ Setup and Installation

To get this project up and running locally, follow these steps: Clone the repository:

Create a virtual environment (recommended): python -m venv venv source venv/bin/activate # On Windows: venv\Scripts\activate

Install dependencies: pip install -r requirements.txt

Set up environment variables: Create a .env file in the root directory of the project and add your API keys:

GOOGLE_API_KEY="your_api_key" LANGSMITH_TRACING=true LANGSMITH_ENDPOINT="https://api.smith.langchain.com" LANGSMITH_API_KEY="your_langsmith_api_key" LANGSMITH_PROJECT="project_name"

• Refer to Google Cloud documentation for obtaining a Gemini API key and its client secrets also obtain client secrets for after enabling the youtub v3 api from the google cloud console add desktop client add in it http://localhost:8080 download the client_secret.json
• The first time you run the upload file it will ask a token which you can generate by logging in provided url by the program, after first time it will create yt_upload.py-oauth2.json to upload in future
• Install ffmpeg: This project relies on ffmpeg for video processing. Download and install it for your operating system from the official FFmpeg website and ensure it's accessible in your system's PATH.
• Install Yt Dlp - Guide to install , after installing add your browser cookies using this extension check here store cookies in cookie.txt in the same folder for full guide about cookies in ytdlp - here

🏃‍♀️ Usage

• Once set up, you can run the main automation script: python agent.py The script will begin its cycle of searching, processing, and uploading. Output messages will be printed to the console indicating the progress.
• For customization you can edits the prompts in prompts.py yourself to get the output for your needs.
• You can change the model name and args inside the agent file to get best out of the llm.

🤖Automation

You can automate this agent to run using the crontab in linux make a shell script for ex:

#!/bin/bash

# Navigate to your project directory
cd ./Agent
# Activate your virtual environment (if applicable)
source venv/bin/activate

# Run your Python script
python agent.py

# Deactivate the virtual environment (optional)
deactivate

chmod +x ./agent.sh

add to crontab

$ crontab -e for example to run every day at 19:30

30 19 * * * /home/ubuntu/agent.sh >> /home/ubuntu/agent_output.log 2>&1

add this to crontab

Screenshots