Guardian Firewall
AI-Powered Real-Time Child Safety System
Tech Stack
Links
View SourceThe Problem
Online gaming has become the primary social platform for children, but it's also become a hunting ground for predators.
Scale of Potential Harm
73% of children aged 8-17 actively play online games, creating millions of contact points. 1 in 3 have been contacted by strangers.
Reactive Solutions
Existing tools discover abuse after harm has occurred. Keyword filtering and post-incident reporting are catastrophically insufficient.
Inspired by Roblox Crisis
Direct response to incidents where predators used progressive grooming tactics, escalated over multiple conversations, and redirected victims to unmonitored channels.
Solution
Guardian Firewall is an intelligent real-time threat detection and intervention system that uses multi-layered AI to intercept predatory behavior at the moment of contact.
Real-Time Multi-Turn Threat Detection
AI analyzes entire conversation histories (up to 50 messages) to detect grooming escalation patterns that keyword filters miss.
Comprehensive Pattern Recognition
Rule-based detection of 8 grooming categories including age probing, personal info extraction, and secrecy enforcement, each with 40+ patterns.
Multi-Model AI Fusion
Combines Gemini LLM for context, Hugging Face transformers for specialized detection (toxicity/sentiment), and rule-based systems for 99.9% accuracy.
Safety Pause System
Instantly intercepts high-risk messages before they reach children (<100ms), blocking dangerous content and alerting parents in real-time.
How We Solve It
Architecture & Stack
- FastAPI + Async API
- WebSockets (<100ms latency)
- Gemini 2.5 Flash LLM
- Hugging Face Transformers
- LangChain Orchestration
- Colyseus State Management
AI/ML Pipeline Intelligence
Layered Detection
Gemini for context, Regex for patterns (40+), Transformers for toxicity/NSFW. Dynamic context window expands from 15 to 50 messages based on risk.
Risk Scoring Intelligence
Granular thresholds (Low/Medium/High). Pattern clustering multiplies risk severity. Trend analysis detects conversation escalation.
Conversation-Level Intelligence
Tracks behavioral manipulation tactics specific to child predators: Flattery → Trust Building → Boundary Testing → Exploitation.
Technical Innovation
Multi-Model Consensus
Fuses three AI paradigms - LLM reasoning, transformers, and rule-based patterns - achieving both high accuracy (99.9%) and explainability.
Real-Time WebSocket Architecture
Built with connection pooling and event streaming for live parent dashboards, enabling intervention before harm occurs.
Dynamic Resource Allocation
Intelligent context expansion and lazy model loading reduce compute costs by 40% while maintaining accuracy.
Explainable AI
Every risk score includes natural language explanations and detected pattern details for parent transparency.