Enterprise-grade, multi-modal AI content detection engine.
Catch ChatGPT essays, Midjourney images, Sora deepfakes, and ElevenLabs voice clones — all in one platform.
AI-generated content is everywhere — and existing detection tools are fragmented. Teachers juggle separate tools for text, images, and video. Students get falsely flagged. Deepfakes go undetected. There is no single platform that handles all modalities.
DetectAi solves this with a unified, multi-modal detection engine that analyzes text, images, video, and audio through a single API and dashboard.
DetectAi uses a Branch-and-Fuse architecture. Each media type is processed through an independent pipeline composed of specialized analysis branches. Branch outputs are fused into a single, calibrated confidence score.
graph TB
subgraph Client["🖥️ Frontend — React + Vite"]
UI[Landing Page / Dashboard]
Auth[Clerk Auth]
SEO[Programmatic SEO Engine]
end
subgraph Gateway["⚡ FastAPI Gateway — Azure App Service"]
direction TB
TextRouter["/detect/text"]
ImageRouter["/detect/image"]
VideoRouter["/detect/video"]
AudioRouter["/detect/audio"]
end
subgraph TextEngine["📝 Text Engine — 5 Branches"]
T1[RoBERTa Transformer]
T2[GPT-2 Perplexity]
T3[Stylometry Analysis]
T4[Sentence-BERT Embeddings]
T5[Humanization Detector]
TF[LightGBM Meta-Classifier]
end
subgraph ImageEngine["🖼️ Image Engine — 6 Branches"]
I1[ViT Deep Classifier]
I2[Error Level Analysis]
I3[FFT Spectral Analysis]
I4[EXIF Metadata Inspection]
I5[PRNU Sensor Noise]
I6[Statistical Heuristics]
IF[Weighted Confidence Fusion]
end
subgraph VideoEngine["🎬 Video Engine — 9 Branches"]
V1[ViT + DinoV2 Ensemble]
V2[Temporal Consistency]
V3[RAFT Optical Flow]
V4[MiDaS Depth Physics]
V5[CLIP Semantic Drift]
V6[FFT Spectral Score]
V7[PRNU Sensor Noise]
V8[Audio-Visual Sync]
V9[Codec Heuristics]
VF[9-Branch Weighted Fusion]
end
subgraph AudioEngine["🎙️ Audio Engine"]
A1[MFCC + Spectral Features]
A2[Calibrated Random Forest]
end
UI -->|HTTPS| Gateway
TextRouter --> TextEngine
ImageRouter --> ImageEngine
VideoRouter --> VideoEngine
AudioRouter --> AudioEngine
T1 & T2 & T3 & T4 & T5 --> TF
I1 & I2 & I3 & I4 & I5 & I6 --> IF
V1 & V2 & V3 & V4 & V5 & V6 & V7 & V8 & V9 --> VF
A1 --> A2
📖 Deep dive: See docs/architecture.md for full technical details on each branch.
| Feature | Description |
|---|---|
| 🔤 Text Detection | 5-branch ensemble (RoBERTa, perplexity, stylometry, embeddings, humanization) with LightGBM meta-classifier and per-sentence explainability |
| 🖼️ Image Detection | 6-branch fusion (ViT, ELA, FFT, EXIF, PRNU, heuristics) with perceptual hash caching |
| 🎬 Video Detection | 9-branch fusion (ViT, DinoV2, RAFT optical flow, MiDaS depth, CLIP semantic drift, temporal analysis) |
| 🎙️ Audio Detection | Spectral feature extraction + calibrated Random Forest for voice clone detection |
| 📊 Dashboard | Real-time scan results, history tracking, analytics charts |
| 🔐 Authentication | Clerk-powered OAuth with SSO support |
| 🌗 Dark Mode | Full dark/light theme support with smooth transitions |
| 🔍 SEO Engine | Programmatic long-tail SEO with Schema.org structured data and auto-generated sitemap |
| Layer | Technology |
|---|---|
| Frontend | React 19, Vite, Framer Motion, Recharts |
| Auth | Clerk |
| Backend | FastAPI, Python 3.11, Uvicorn |
| ML Models | PyTorch, HuggingFace Transformers (ViT, RoBERTa, DinoV2, RAFT, MiDaS, GPT-2) |
| Ensemble | LightGBM, scikit-learn (isotonic calibration) |
| Task Queue | Celery + Redis |
| Database | SQLite |
| Hosting | Azure App Service (B3), Vercel |
| CI/CD | GitHub Actions → Azure auto-deploy |
| Analytics | Vercel Web Analytics |
Click to expand directory tree
detectai/
├── backend/ # FastAPI application
│ ├── main.py # App entrypoint & router registration
│ ├── text/ # Text detection module
│ │ ├── router.py # /detect/text endpoint
│ │ ├── engines/ # 5 detection branches
│ │ │ ├── transformer.py # └─ RoBERTa classifier
│ │ │ ├── perplexity.py # └─ GPT-2 perplexity scorer
│ │ │ ├── stylometry.py # └─ Statistical text analysis
│ │ │ ├── embeddings.py # └─ Sentence-BERT similarity
│ │ │ └── humanization.py # └─ Paraphrase detector
│ │ ├── ensemble/ # Meta-classifier & calibration
│ │ ├── explainability/ # Per-sentence attribution
│ │ ├── preprocessors/ # Chunking & language detection
│ │ └── models/ # Serialized model artifacts
│ ├── image/ # Image detection module
│ │ ├── router.py # /detect/image endpoint
│ │ ├── fusion.py # Weighted branch fusion
│ │ ├── tasks.py # Celery async tasks
│ │ └── branches/ # 6 detection branches
│ │ ├── deep_classifier.py # ViT classifier
│ │ ├── ela.py # Error Level Analysis
│ │ ├── fft_cnn.py # FFT spectral analysis
│ │ ├── metadata.py # EXIF inspection
│ │ ├── prnu_residual.py # Sensor noise analysis
│ │ └── advanced_heuristics.py
│ ├── video/ # Video detection module
│ │ ├── router.py # /detect/video endpoint
│ │ └── branches/ # 9 detection branches
│ │ ├── vit_ensemble.py # ViT + DinoV2
│ │ ├── temporal_consistency.py
│ │ ├── motion_physics.py # RAFT optical flow
│ │ ├── depth_physics.py # MiDaS depth estimation
│ │ ├── semantic_drift.py # CLIP embeddings
│ │ ├── fft_score.py
│ │ ├── prnu_score.py
│ │ ├── audio_physics.py
│ │ └── advanced_heuristics.py
│ ├── audio/ # Audio detection module
│ │ ├── router.py # /detect/audio endpoint
│ │ └── tasks.py # Feature extraction & classification
│ ├── database.py # SQLite connection manager
│ ├── history.py # Scan history CRUD
│ ├── models.py # Pydantic schemas
│ ├── celery_app.py # Task queue config
│ ├── Dockerfile # Container build
│ └── requirements.txt # Python dependencies
├── frontend/ # React SPA
│ ├── src/
│ │ ├── App.jsx # Router & providers
│ │ ├── pages/ # Page components
│ │ ├── components/ # Shared UI components
│ │ ├── config/seoConfig.js # Programmatic SEO configuration
│ │ ├── context/ # Theme provider
│ │ └── lib/ # Utility functions
│ ├── public/ # Static assets, sitemap, robots.txt
│ ├── index.html # HTML entry point
│ └── vite.config.js # Vite build configuration
├── docs/ # Technical documentation
│ ├── architecture.md # System architecture deep-dive
│ ├── api_reference.md # API endpoint reference
│ └── text_engine_architecture.md
├── .github/workflows/ # CI/CD pipelines
│ └── azure-deploy.yml # Auto-deploy to Azure on push
├── .gitignore
├── CONTRIBUTING.md
├── LICENSE # All Rights Reserved
└── README.md # ← You are here
- Python 3.11+
- Node.js 18+
- Git
git clone https://github.com/Harsh2o/detectai.git
cd detectaicd backend
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env # Add your HuggingFace token
uvicorn main:app --reload --host 0.0.0.0 --port 8000The API will be live at http://localhost:8000/docs
cd frontend
npm install
cp .env.example .env.local # Configure API URL & Clerk keys
npm run devThe app will be live at http://localhost:5173
The backend auto-deploys via GitHub Actions. Every push to main triggers the workflow in .github/workflows/azure-deploy.yml.
# Deploys to: https://detectai-api-xxxxx.azurewebsites.netThe frontend auto-deploys via Vercel's GitHub integration. Connect your repo in the Vercel dashboard and set the following environment variables:
| Variable | Value |
|---|---|
VITE_API_URL |
Your Azure backend URL |
VITE_CLERK_PUBLISHABLE_KEY |
Your Clerk publishable key |
| Document | Description |
|---|---|
| Architecture | Full system architecture, engine details, and infrastructure |
| API Reference | Complete REST API endpoint documentation |
| Text Engine | Deep-dive into the text detection pipeline |
| Contributing | Development setup and contribution guidelines |
This project is licensed under All Rights Reserved — see the LICENSE file for details. You may view this code for educational purposes, but copying, deploying, or modifying it is strictly prohibited.
Built with ❤️ by Harsh Gupta