FoodChooser - AI Nutrition Analyzer

When choosing dishes at restaurants or cafes, it's difficult to quickly assess their nutritional value and alignment with body recomposition goals. Traditional calorie calculators require manual data entry, and existing apps don't account for personal restrictions or provide practical optimization advice at the point of ordering. We developed an intelligent nutrition analysis system based on a multi-agent architecture with four specialized AI agents. The Router Agent classifies image type (menu or prepared dish), determines meal time, and selects the appropriate agent for processing. The Menu Advisor Agent helps choose optimal dishes from menus considering blacklisted foods (pork, beef, fried foods, desserts). The key feature is the Dish Scorer Agent, which analyzes photos of prepared dishes using a 100-point scoring system. The scoring is based on four metrics: protein density (0-35 points), satiety index accounting for fiber and glycemic index (0-30 points), anti-bloating score for digestive comfort (0-25 points), and macrobalance (0-10 points). The system adapts recommendations based on time of day - breakfast allows up to 50% carbs, while dinner prioritizes protein and vegetables. For dishes scoring 60-79 points, the Optimizer Agent automatically activates to suggest practical modifications: protein boost (add eggs, chicken), volume therapy (more vegetables for satiety), anti-bloating (remove gas-producing foods), and macrobalance. Each recommendation includes specific instructions for waiters and predicted score improvement after changes. The technical implementation is built on n8n workflow automation with Google Gemini 2.5 Flash Lite integration for the Router Agent (temperature 0 for determinism) and standard Gemini for other agents. We implemented Structured Output Parser for JSON schema validation, ensuring reliable data outputs. HTML-safe text generation protects against Telegram parsing errors, while API cost optimization is achieved through temperature control (0.1 for Dish Scorer, 0.7 for creative Optimizer). The system uses computer vision for visual assessment of portions, cooking methods, and ingredients without connecting to external nutritional databases. All analysis completes in 20-30 seconds with detailed breakdown of each scoring component, educational explanations of body composition impact, and scientific justification for recommendations.