🛡️ Minor-Guard — Age-Aware AI Safety System

A child-safe AI system that combines Retrieval-Augmented Generation (RAG) with age-aware guardrails to deliver safe, accurate, and developmentally appropriate responses for users aged 5–17.

---

🚀 Overview

This project implements a defense-in-depth architecture for AI safety. Unlike standard chatbots, this system ensures that responses are Safe + Useful + Age-Appropriate simultaneously.

It achieves this through:

• 🔍 RAG (Retrieval-Augmented Generation): Grounds answers in verified information to prevent hallucinations.
• 🧠 Age-aware Guardrails: Tailors the complexity, tone, and safety constraints based on the user's developmental stage.
• 🛡️ Adversarial Resistance: Actively detects and blocks harmful or jailbreak prompts.
• 📊 Evaluation Engine: Systematically measures safety, accuracy, and performance across different configurations.

---

🧱 Architecture Workflow

The system processes requests through a multi-stage pipeline to ensure safety and quality:

User prompt + age ➡️ Topic Router ➡️ RAG Context Retrieval ➡️ Guardrails + Logic ➡️ Safe Response

---

📁 Project Structure

AI_Project/
├── app.py                  # Flask web application & API endpoints
├── main.py                 # CLI interface for chat, evaluation, and charts
├── simulator.py            # Core simulation and testing logic
├── requirements.txt        # Python dependencies
├── README.md               # Project documentation
├── data/
│   └── prompts.py          # System and test prompts
├── rag/
│   ├── database.py         # Vector/knowledge database integration
│   └── metaprompts.py      # Meta-level instructions for RAG
├── evaluation/
│   └── evaluator.py        # Scoring and metric computation
└── visualization/
    └── charts.py           # Generation of performance graphs

---

⚙️ Setup & Installation

Local Environment

1. Clone the repository and navigate to the project directory:

   cd AI_Project

2. Install the required dependencies:

   pip install -r requirements.txt

---

💬 Usage

The project provides a comprehensive CLI (main.py) for various operations:

1. Interactive Chat

Run a simulated chat session with age-specific parameters:

python main.py --mode chat --age 12 --prompt "what is depression"

2. Run Evaluations

Execute the evaluation suite across different system configurations:

python main.py --mode eval

3. Generate Visualizations

Create charts comparing system performances:

python main.py --mode charts --demo

---

🌐 API Endpoints (Cloud Run)

The application is deployed on Google Cloud Run and exposed via a REST API.

Base URL:
https://ai-project-b5dl-d5pfcfjqf-aniyastiks-projects.vercel.app/

Endpoints

• Health Check

  • GET /health
  • Returns the operational status of the API.

• Chat Interaction

  • POST /chat
  • Body Example:

    {
      "prompt": "what is depression",
      "age": 12,
      "system": "Proposed"
    }

• Run Remote Evaluation

  • POST /evaluate

---

🧪 Systems Compared

The evaluation engine tests our Proposed architecture against standard approaches:

1. Baseline: Standard LLM (No RAG, No Guardrails). Prone to hallucinations and unsafe content.
2. RAG Only: Grounded in facts but lacks safety filtering for sensitive topics.
3. Guardrails Only: Safe and filtered but not factually grounded (can still hallucinate safe responses).
4. Proposed (Full System): Combines RAG and Guardrails for optimal performance.

---

📊 Evaluation Metrics & Results

The system is evaluated on four key pillars:

• Safety: Prevention of harmful content.
• Accuracy: Factual correctness of the response.
• Hallucination Resistance: Ability to avoid making up facts.
• Age Alignment: Appropriateness of the language and concepts for the target age.

Results Summary

System	Safety Score	Accuracy Score
Baseline	0.60	0.00
RAG Only	1.00	0.72
Guardrails Only	1.00	1.00
Proposed (Full System)	0.98	0.98

---

🚀 Deployment

• Frontend: Deployed on Vercel
• Infrastructure: Google Cloud Run
• Pipeline: GitHub CI/CD

---

👥 Authors

Aniya Baghirova, Shehana Byramli, Parviz Bayramli