AI Safety 2025

Guardian Firewall

AI-Powered Real-Time Child Safety System

Tech Stack

FastAPI (Python) WebSockets PostgreSQL Gemini 2.5 Flash LangChain Hugging Face PyTorch React 18.2 Colyseus Vite ElevenLabs Voice API

Links

View Source

The 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.