NUTRIVISION AI
Indonesian Food Image Classification & LLM Nutrition Recommendation
PROJECT BACKGROUND
Penelitian skripsi S1 Teknik Informatika di Universitas Gunadarma yang berfokus pada pengenalan kuliner lokal dan edukasi nutrisi berbasis kecerdasan buatan.
THE CORE CHALLENGE
Kuliner nusantara memiliki keberagaman visual yang sangat tinggi (kuah gulai, sambal, potongan rempah, variasi plating) yang sulit diklasifikasikan oleh model umum.
ENGINEERING STRATEGY
Mengumpulkan dan membersihkan dataset 8.844 citra makanan lokal, menerapkan transfer learning MobileNetV2 dengan augmentasi data, dan menghubungkan hasil klasifikasi ke prompt engineering LLM terstruktur.
ARCHITECTURE SPECIFICATION
- 01Dataset Pipeline: 8,844 images preprocessed with adaptive resizing, augmentation, and normalization
- 02Inference Model: MobileNetV2 with fine-tuned Dense layers and Dropout 0.3 for multi-class prediction
- 03Reasoning Engine: DeepSeek R1 prompt chain generating macronutrient analysis and dietary tips
- 04Web Service: Flask REST API served with Gunicorn WSGI and Nginx on Linux Ubuntu VPS
Dataset Preprocess
8,844 images resized (224x224), augmented with rotation/zoom, and normalized.
MobileNetV2 CNN
Transfer learning feature extraction + Dense Dropout (0.3) Softmax classification.
DeepSeek R1 Chain
Prompt engineering chain producing caloric breakdown and dietary recommendations.
Flask VPS API
Served via Gunicorn WSGI server and Nginx reverse proxy on Linux Ubuntu.
DELIVERED CAPABILITIES
- Multi-class visual classification across signature Indonesian traditional dishes
- Automated estimation of portion size and macronutrient caloric breakdown
- Interactive dietary recommendations tailored to detected meal components
- Lightweight inference engine capable of running efficiently on affordable cloud VPS
PROJECT LIFECYCLE
Research prototype and academic publication codebase.
TECHNICAL TAKEAWAY
Menggabungkan visi komputer yang presisi dengan kemampuan penalaran LLM menghasilkan aplikasi edukasi gizi yang jauh lebih kontekstual dan bermanfaat.
