---

🛡️ SafeScroll — AI-Powered Social Media Safety Auditor

Making your social feed safer — one scroll at a time.

SafeScroll is an AI-powered Trust & Safety auditing system that simulates how modern social media platforms internally detect and respond to harmful content and unsafe user behavior.

It demonstrates how multi-agent GenAI systems can be applied to:

• content moderation
• underage safety
• grooming detection
• policy enforcement
• human-readable safety reporting

This project is built for research, demos, and resume-ready showcasing of real-world AI safety engineering.

---

🧠 How SafeScroll Works

SafeScroll uses a multi-agent GenAI architecture, where each agent focuses on a specific safety responsibility.

🔹 1. Underage Risk Agent

• Evaluates declared age vs. content style
• Flags possible underage misrepresentation
• Outputs a risk score with reasoning

🔹 2. Content Safety Agent

Analyzes posts for:

• Bullying & harassment
• Self-harm indicators
• Sexual exploitation cues
• Substance abuse references

Produces both per-post and aggregated risk levels.

🔹 3. Interaction (DM) Risk Agent

• Detects grooming patterns
• Identifies secrecy pressure & power imbalance
• Considers age differences in conversations

🔹 4. Policy Violation Agent

• Maps AI findings to company safety policies
• Determines severity
• Recommends moderation actions (monitor, warn, restrict, escalate)

🔹 5. Safety Report Generator

• Combines all agent outputs
• Generates a clear, human-readable safety report
• Summarizes risks, evidence, and final recommendation

---

🛠️ Tech Stack

Layer	Technology
UI	Streamlit
Language	Python
AI / LLM	OpenAI GPT (multi-agent setup)
Data	Synthetic CSV datasets
Architecture	Multi-Agent Safety System
Styling	Custom Streamlit theme
Charts	Streamlit native charts

---

🗂 Folder structure

 safe-scroll/
├── .streamlit/
│   └── config.toml
├── assets/
│   ├── safescroll_overview.png
│   └── safescroll_moderation.png
├── data/
│   ├── interactions.csv
│   ├── posts.csv
│   └── users.csv
├── policies/
│   └── safety_policies.txt
├── README.md
├── agents.py
├── app.py
├── generate_synthetic_data.py
└── requirements.txt

---

▶️ Running the App Locally

1️⃣ Clone the repository

git clone https://github.com/nirjanadas/safe-scroll.git
cd safe-scroll

2️⃣ Install dependencies

pip install -r requirements.txt

3️⃣ Run the app

streamlit run app.py

---

📊 Output & Reports

SafeScroll generates:

• Structured JSON risk analysis
• Visual risk indicators
• Policy violation summary
• Final human-readable safety report

---

📌 Disclaimer

All user data is synthetic and used only for research and demonstration purposes.