AI Director

A fully localized, serverless video ingestion pipeline and conversational AI engine optimized explicitly for Apple Silicon (M-series Macs).

AI Director allows you to dump entire SD cards or massive folders of raw video clips (.mp4, .mov, .avi, .mkv, etc.) onto a drive. The engine will autonomously scan them, generate proxy thumbnails, transcribe all dialogue locally using MLX Whisper, and generate semantic visual tags using a local Vision Language Model (MLX LLaVA).

You can then search, filter, and chat with your entire video archive via a beautiful Next.js frontend, or ask the AI to generate JSON timelines to automatically import into Premiere Pro/DaVinci Resolve! The Studio now features a powerful "Hide Selected" tool to quickly mark redundant clips as duplicates.

System Requirements

This project is explicitly tailored for high-performance execution on Mac hardware using Apple's MLX framework.

• Hardware: Apple Silicon (M1, M2, M3, M4)
• OS: macOS 13.3+ minimum
• External Dependencies: ffmpeg and ffprobe must be installed on your system.

  brew install ffmpeg

🚀 Quick Start (One-Click Launch)

This project includes a smart startup script that automatically handles all Python virtual environments, pip dependencies, and npm Node modules for you on its first run.

To install and run the application: Simply double-click the Launch AIDirector.command file in the project directory.

Alternatively, you can run it from the terminal:

./Launch\ AIDirector.command

Note: On the very first run, it may take 2-3 minutes to download all necessary dependencies. Once complete, it will automatically spin up the FastAPI background process and launch the Next.js UI.

Finally, open your browser to http://localhost:3000.

Initial Configuration

1. Click the Settings (⚙️) tab in the sidebar.
2. Enter the absolute path where you want the SQLite database to live (e.g., /Users/yourname/Desktop/Video_Archive.db).
3. Add the absolute folders mapping to your raw video files under Target Scan Directories.
4. Enter your Gemini API key (this is passed to the LLM agent that handles natural language searches—your actual video transcription and tagging is all done 100% locally).
5. Click Scan Directory & Process Videos to start building your database!
6. Use the Archive Maintenance section to bulk-hide duplicates (proximity-based) or short Live Photos (< 3s).

Note on First Run: The exact first time you click "Scan", the backend will automatically pull the MLX versions of Whisper and LLaVA from HuggingFace. This initial download can take several minutes and occupy 5-10GB of disk space.

Built by Waylon Kenning.