What it does

This repo is a cookbook for building a Facebook Messenger bot that replies like you. It scrapes your conversation history from Facebook, Google Hangouts, LinkedIn, Discord, and WhatsApp, cleans it into (message, response) pairs, trains a seq2seq model in TensorFlow 1.x, and wires the result to a Flask backend that a Node.js Messenger bot calls for replies.

The interesting bit

The project treats your entire digital social footprint as training corpus—group chats, DMs, the lot—and the README is essentially a field guide to wrangling each platform’s idiosyncratic export format. The WhatsApp instructions alone involve emailing yourself a chat log and running it through a third-party parser that may or may not still exist.

Key highlights

• Supports five chat platforms with custom parsers for each export format
• Joint training option: skip Word2Vec pre-training and let seq2seq learn embeddings itself
• Saves intermediate numpy matrices at each stage so you can iterate without reprocessing
• Includes a separate Flask server repo for model deployment
• TensorFlow 1.0+ era code, with all the checkpoint-file archaeology that implies

Caveats

• TensorFlow 1.x dependency; the README does not mention migration paths
• WhatsApp parser link is dead; fallback to manual .txt formatting with specific date conventions
• No evaluation metrics or loss curves shown; “interesting responses that (hopefully) resemble yourself” is the stated success criterion

Verdict

Worth a weekend if you’ve ever wanted to automate your own small talk and don’t mind debugging decade-old TF APIs. Skip it if you need production reliability or modern framework support.