PixelML/epstein-files-cctv-video-memory-search

Epstein Files CCTV — Video Memory Search

Search 472 AI-generated temporal event captions + 21 structured summaries & analysis reports from DOJ Epstein Files Dataset 8 (MCC prison CCTV surveillance footage).

GPT-4.1 vision analyzed 30-second video chunks to describe what happens across frames — temporal changes, people entering/leaving, door activity. Static scenes are filtered out.

Built with `av` — video memory for AI agents.

How it works

1. 1.25 CCTV clips from MCC New York (July–August 2019) were ingested with av ingest --captions --topic security
2. 2.Three-layer cascade: chunk VLM (GPT-4.1 vision) → structured event log → analysis report
3. 3.472 temporal event captions + 21 summaries + 21 reports indexed with SQLite FTS5
4. 4.This Space lets you search those captions with natural language

Links

• —Dataset
• —av CLI
• —agentic.video