Building 1,000+ AI Personas for Telegram Comments

The article describes how the authors built an AI system to generate contextual comments from 1,000+ unique personas on Telegram channels, overcoming the challenges of Telegram's anti-bot systems and building a 15-layer anti-spam pipeline.

Details

The article describes how the authors tried hiring human teams to write comments on Telegram channels, but it failed after 3 months due to issues like boredom, inconsistency, and no-shows. They then built an AI system using Telegram's official C++ library, TDLib, to generate comments from over 1,000 unique personas. Each persona has persistent personality traits, opinion drift over time, and natural language quirks like typos and slang. Around 65-85% of the generated comments are threaded replies where the personas argue, agree, and reference real-time market data from services like BingX, CoinGlass, and CoinGecko. The biggest technical challenge was not the natural language processing, but keeping the accounts alive and avoiding Telegram's aggressive anti-bot systems. They had to implement natural typing delays, randomized activity patterns, and session management on TDLib to avoid bans. The authors also built a 15-layer anti-spam pipeline called ModerAI with 99.7% accuracy, including voice spam transcription and Vision AI for image spam.