DeshiPlate AI is an intelligent food-recognition and nutrition-guidance system built specifically for Bangladeshi cuisine.
It combines a NextViT-based image classifier, a Next.js web app, and an LLM-powered nutrition engine to help users understand their meals, track macros, and receive personalized dietary recommendations.
🔗 Live Site: deshiplateai.vercel.app
- Custom dataset with 33 popular Bangladeshi foods
- Fine-tuned NextViT model
- Hosted on Hugging Face Hub
- Provides label + confidence + nutrition info
- Calculates:
- Daily caloric needs
- Recommended protein, carbs, fat
- BMI
- Tracks daily meals
- Performs food health suitability checks based on:
- Diabetes
- BP
- High cholesterol
- Weight goals
- Activity level
- Generates personalized meal advice
- Identifies risky foods
- Suggests alternatives
- Explains results in natural language
- Next.js frontend
- Drizzle ORM + PostgreSQL
- NextAuth authentication
- TailwindCSS UI
- HuggingFace inference API
DeshiPlate-AI/
│
├── frontend/ # Production-grade Next.js web application
├── model-training/ # NextViT training pipeline for 33-class food classifier
└── README.md
- User uploads food image
- Frontend calls
/api/food/analyze - API forwards image to HuggingFace model
- Model returns class + confidence
- System fetches nutrition information
- LLM generates personalized guidance
- User gets full breakdown + suggestions
| Layer | Technologies |
|---|---|
| Model Training | Python, PyTorch, NextViT |
| Model Hosting | HuggingFace Hub |
| Frontend | Next.js 16, TypeScript, React |
| Backend / API | Next.js Server Actions, Route Handlers |
| Database | PostgreSQL + Drizzle ORM |
| Authentication | NextAuth |
| Styling | TailwindCSS |
| AI Assistant | OpenAI LLM / ChatGPT API |
Released under the MIT License.
If DeshiPlate AI helps you, please consider starring the repo to support future development.
Contributions are always welcome:
- Add new Bangladeshi foods
- Improve accuracy
- Extend nutrition database
- UX/UI upgrades