SentinelAI

AI Powered Hate Speech Detection using BiLSTM + Attention

Live Demo → sentinelai-egh5.onrender.com

SentinelAI is a deep learning system that detects hate speech and toxic content in real time. The project uses a custom Bidirectional LSTM with Multi-Head Self-Attention model — trained from scratch on 56,000+ samples — integrated with a FastAPI web application and a modern dark-theme interface.

---

Overview

Online platforms generate massive amounts of user generated content every day. Detecting abusive or hateful language automatically is critical for moderation and safe communication.

SentinelAI provides an AI powered system that:

• Detects hate speech, toxicity, and offensive language
• Uses a custom BiLSTM + Attention deep learning model (no pretrained weights, no API calls)
• Provides real time predictions with a toxicity score and severity classification
• Delivers results through a professional dark-theme web interface with animated visualizations
• Can be deployed on any container platform (Docker ready)

---

Model

The project uses a custom BiLSTM + Multi-Head Self-Attention model trained from scratch — no HuggingFace, no pretrained weights, no external API.

Property	Value
Architecture	Embedding → BiLSTM → Multi-Head Attention → LayerNorm → FC
Parameters	~3.1M
Vocabulary	20,000 words (custom WordTokenizer, serialized to JSON)
Training data	56,701 samples (raw_data.csv + imbalanced_data.csv)
Loss	BCEWithLogitsLoss with pos_weight (handles class imbalance)
Optimizer	AdamW + ReduceLROnPlateau scheduler
Accuracy	95%
F1 Score	0.93
Classes	0 → No Hate, 1 → Hate / Abusive

Model artifacts (model.pt, vocab.json, config.json) are committed to the repository — zero downloads at runtime.

---

Features

• Real time hate speech and toxicity detection
• Custom BiLSTM + Multi-Head Self-Attention (trained from scratch)
• Toxicity score with animated count-up display
• Green → Yellow → Red spectrum bar with position marker
• Severity classification: CLEAR / LOW / MEDIUM / HIGH / CRITICAL
• Animated neural network canvas background
• Gradient shield logo and SVG favicon
• Two-column result layout (score + breakdown)
• Lazy model loading on first request
• Docker ready, deployed on Render

---

Tech Stack

Machine Learning

• Python 3.11
• PyTorch (CPU)
• Custom WordTokenizer (word-level, JSON serialized)
• Scikit-learn (train/test split, metrics)
• Pandas / NumPy

Backend

• FastAPI
• Uvicorn / Gunicorn
• Jinja2 (server-side rendering)

Frontend

• HTML / CSS / JavaScript
• Inter + JetBrains Mono (Google Fonts)
• Canvas-based animated neural network background
• No CSS framework (custom design system)

---

Project Structure

SentinelAI/
│
├── app.py                        # FastAPI application entry point
│
├── routers/
│   └── predict.py                # GET / and POST /predict routes
│
├── services/
│   ├── classifier.py             # BiLSTM model + WordTokenizer definitions
│   └── inference.py              # Lazy model loading + predict_text()
│
├── training/
│   └── train_model.py            # Standalone training script
│
├── templates/
│   ├── base.html                 # Layout: topbar, canvas, fonts
│   └── index.html                # Input form + result card
│
├── static/
│   ├── css/styles.css
│   ├── js/script.js
│   └── favicon.svg
│
├── artifacts/
│   └── model/
│       ├── model.pt              # Trained model weights (12 MB)
│       ├── vocab.json            # WordTokenizer vocabulary
│       └── config.json           # Model hyperparameters
│
├── data/
│   ├── raw_data.csv
│   └── imbalanced_data.csv
│
├── Dockerfile
├── render.yaml
├── requirements.txt
└── runtime.txt

---

Installation

git clone https://github.com/Naman21036/SentinelAI.git
cd SentinelAI

python -m venv venv
# Windows
venv\Scripts\activate
# Linux / Mac
source venv/bin/activate

pip install -r requirements.txt

---

Running the Application

python -m uvicorn app:app --reload --port 8000

Open in browser: http://127.0.0.1:8000

---

Training the Model

If you want to retrain from scratch:

python -m training.train_model

This reads data/raw_data.csv and data/imbalanced_data.csv, trains the BiLSTM model, and saves artifacts to artifacts/model/.

---

Deployment

The app is Docker ready. It uses the CPU-only PyTorch wheel to keep the image size small.

docker build -t sentinelai .
docker run -p 8000:8000 sentinelai

Render: Connect the GitHub repo. Render uses render.yaml to detect Docker environment automatically. Set PYTHON_VERSION=3.11.9 as an environment variable if using the native Python runtime.

---

Example

Input

You are a useless idiot

Output

Threat Detected · HIGH
Toxicity Score: 82.4%
Safe: 17.6%  |  Toxic: 82.4%

---

Use Cases

• Content moderation systems
• Social media platforms
• Online communities
• Gaming chat monitoring
• Comment filtering pipelines

---

Contributing

Contributions are welcome.

1. Fork the repository
2. Create a feature branch
3. Submit a pull request

---

License

This project is licensed under the MIT License.

---

Author

Naman Gupta
BIT Mesra
AI Enginnering Enthusiast