• 

  Harmful content detected

• 

  content appears safe

🛡️ About the Project – SafeTweet: AI Bot Against Cyberbullying & Hate Speech 💡 Inspiration In the age of social media, platforms like Twitter/X are flooded with cyberbullying, hate speech, and misinformation. This creates a toxic environment—especially harmful to youth and marginalized communities. We were inspired to build a tool that promotes digital well-being and positive online interaction.

That’s how SafeTweet was born: A bot that automatically detects harmful tweets, finds factual help using RAG (Retrieval-Augmented Generation), and responds kindly to foster healthier conversations.

🛠️ How We Built It We developed the project using an open-source and transparent ML pipeline:

Frontend: Streamlit app, deployed via Hugging Face Spaces

Backend: Python Flask backend (lightweight & interpretable)

ML Model: Custom-trained scikit-learn text classification model

RAG System: Semantic search using FAISS to retrieve guidance/resources

LLM (via Groq): Quantized open-source Mistral model for response generation

Twitter API: Automatically reads & replies to tweets when in bot mode

python Copy Edit

Example: Toxic Tweet Detection

prediction = model.predict(["You are worthless and nobody cares about you."]) print(prediction) # ['toxic'] 🧠 What We Learned Training a text classifier with annotated datasets (non-black-box)

Semantic vector search using FAISS

Deployment via Hugging Face + Streamlit UI customization

Secure integration of the Twitter/X API

Fine-tuning prompt engineering for Groq-hosted LLMs

🚧 Challenges We Faced Collecting high-quality non-LLM datasets for toxic tweet training

Keeping the app responsive on a free-tier deployment

Balancing accuracy with interpretability

Creating respectful auto-responses without sounding robotic

Avoiding false positives and unintended censorship

📌 Summary SafeTweet is an open-source, fully-transparent AI bot that flags harmful tweets and responds with empathetic, fact-based, and resourceful replies.

Built With

• faiss
• flask
• generation)
• groq
• python
• retrieval-augmented
• scikit-learn
• streamlit
• twitter