A software engineer and lodge owner developed a local-first AI indexing tool to make massive, unorganized video archives searchable using natural language queries on a standard laptop.

Key Points

• The tool, named Framedex, uses Claude Code and local LLMs like Gemma 4 31B to generate descriptive metadata sidecars for video files.
• The pipeline integrates WhisperX for transcription, InsightFace for facial recognition, and ExifTool for GPS data to create a queryable archive.
• By building a local index, the developer reduced the cost of managing video assets from $140 per month in SaaS fees to nearly zero.
• The system runs on a 2021 M1 Max MacBook Pro, utilizing Apple’s swap memory to process large language models locally without cloud dependencies.
• The project is open-source and available on GitHub, designed to solve the "unlabeled archive" problem before applying AI-driven video editing.

Why it Matters

This approach highlights a shift in AI utility, moving away from expensive, high-level SaaS editors toward local, infrastructure-level indexing that provides actual data control. By prioritizing a searchable archive, creators can leverage AI to manage massive media libraries efficiently without sacrificing privacy or relying on costly cloud subscriptions.