Story

This project reconstructs the Epstein email records from the recent U.S. House Oversight Committee releases using only public-domain documents (23,124 image files + 2,800 OCR text files).

Most email pages contain only one real message, buried under layers of repeated headers/footers. I wanted to rebuild the conversations without all the surrounding noise.

I used an OCR + vision-LLM pipeline to extract individual messages from the email screenshots, normalize senders/recipients, rebuild timestamps, detect duplicates, and map threads. The output is a structured SQLite database that runs client-side via SQL.js (WebAssembly).

The repository includes the full extraction pipeline, data cleaning scripts, schema, limitations, and implementation notes. The interface is a lightweight PWA that displays the reconstructed messages in a phone-style UI, with links back to every original source image for verification.

Live demo: https://epsteinsphone.org

All source data is from the official public releases; no leaks or private material.

Happy to answer questions about the pipeline, LLM extraction, threading logic, or the PWA implementation.