💡 Inspiration: Hacking Decision Paralysis The inspiration for Foodyzer stems from a critical, high-stakes problem: nutritional decision paralysis. For millions globally—including the 422 million with diabetes and the 500 million avoiding allergens—choosing food is not a simple task; it’s a constant, high-risk health calculation. Existing nutrition apps fail this audience because they are generic, offering simple "A-F" scores that ignore the user's unique medical context (e.g., "Good for a healthy person, but deadly for someone with a specific allergy").

We were inspired to build the last mile solution—a system that delivers immediate, hyper-personalized, context-aware intelligence right at the point of decision, whether scanning a barcode in a store or viewing a restaurant menu.

🛠️ How We Built It: The Multi-Agent Co-Pilot Our core technological achievement is a Multi-Agent Architecture running on the Gemini 2.5 Flash engine. This is the foundation for our "Nutrition Co-Pilot," enabling sophisticated, contextual reasoning far beyond simple API calls.

Key Technical Components:

The Agent System (The Core): We designed five specialized agents that collaborate to process information and deliver the Personalized Suitability Score:

Profile Agent: Manages and contextualizes the user's deep health profile (allergies, chronic conditions, goals).

Vision Agent: Handles multi-modal inputs, analyzing images of ingredient labels and complex, unstructured data like restaurant menus.

Analysis Agent: Performs a dynamic evaluation of the food against the user's profile and returns the final suitability score.

Chat Agent: Powers the Nutri-Chat, managing complex, multi-conditional conversational queries.

External Data Agent: Fetches real-time market data (e.g., price, stock) and scientific information.

Platform: The application is built on React Native / Expo, ensuring a stable, production-ready codebase for rapid deployment across both iOS and Android platforms.

🧠 What We Learned: The Power of Context Our main learning was the absolute necessity of contextual reasoning for hyper-personalization. We realized that simply using an LLM to read a label is insufficient; the model must be structured as an Agent that first consults the Profile Agent before performing any analysis.

We learned that:

Suitability > Scoring: Providing a dynamic score tailored to the individual is far more valuable than a static, generic health score.

Multi-Modal Integration is Key: Seamlessly transitioning between barcode scan, menu photo analysis, and complex chat within one coherent system is what delivers true utility.

🛑 Challenges We Faced Latency Management: The primary challenge was achieving real-time, production-ready performance for the scanner. Coordinating the five agents to deliver a complex, context-aware suitability score in less than two seconds required rigorous optimization and efficient API usage with Gemini 2.5 Flash.

Data Integration and Compliance: Handling sensitive medical data (e.g., allergies, chronic conditions) required meticulous planning to ensure the application is HIPAA/GDPR aligned from day one, which added significant complexity to the Profile Agent's data handling protocols.

UI/UX for Complexity: Translating a complex Multi-Agent architecture into a simple, single, actionable score for the end-user was a continuous design challenge that we solved with the clear, color-coded "Suitability Score" interface.

Built With

Share this project:

Updates

Submission history