← All Projects
In Development

Social Virality System

A self-improving, multi-agent AI system designed to achieve viral dominance on Twitter and MoltBook through data-driven content optimization, ML-powered prediction, and autonomous decision-making.

The Problem

Going viral on social platforms is mostly luck — or it's years of trial and error. What if you could approach it scientifically? Scrape what works, train ML models on viral content, predict performance before posting, and continuously learn from every result?

What I Built

A 5-layer autonomous system: data collection scrapers feed a central SQLite knowledge base, which trains ML models (XGBoost virality predictor trained on 200+ platform posts), powering an autonomous decision engine with human-override safety controls. The system supports multiple AI agents with shared intelligence — each agent has a unique personality but they all learn from each other's data. Extends to Twitter with external intelligence monitoring.

5-Layer Architecture

L5Human InterfaceDaily reports · Discord alerts · Approval system
L4Autonomous DecisionsDecision engine · Risk manager · Flag system
L3IntelligenceVirality predictor · Timing optimizer · Persona optimizer · Pattern discovery
L2Knowledge BaseCentral DB · Learnings · Experiments · Shared memory
L1Data CollectionMoltBook scrapers · Twitter monitor · Reddit monitor · Engagement tracker

Tools & Stack

PythonXGBoostSQLiteWeb ScrapingOpenClawDiscord WebhooksCron JobsMulti-Agent AI

ML Models

  • Virality Predictor: XGBoost — predicts upvotes from 16 content features
  • Engagement Predictor: Random Forest — comment quality & velocity
  • Response Classifier: Logistic Regression — reply priority scoring
  • Content Generator: LLM fine-tuned on top-performing posts
  • Trend Forecaster: Time series — predicts topic trajectory

Challenges & Lessons

The key insight was building compound learning loops: every post generates training data that improves the next prediction. The multi-agent architecture means Agent A's success teaches Agent B what works. Safety guardrails were critical — the flag system pauses autonomous actions when confidence drops below 40% or risk exceeds tolerance. The system's recursive self-improvement (questioning its own question-generation process) is the most intellectually interesting part of the design.

← PreviousPrediction Bots