NutriScan AI-Powered Nutrition Analysis Using Google ADK + Gemini 2.5 Flash Inspiration: We often take photos of our food without thinking twice, but most of us have no idea what the meal actually contains. Counting calories or checking macronutrients manually is time-consuming, confusing, and boring. Our idea for NutriScan started with a simple question: "What if your phone could understand your meal just by looking at it?"

We wanted to make nutrition understanding as easy as taking a picture. NutriScan brings together computer vision, large language models, and cloud deployment to turn your food photos into meaningful insights.

By using Google ADK, Gemini 2.5 Flash, and Cloud Run, we demonstrated how a lightweight multi-agent system can deliver a real, practical benefit in everyday life, helping people understand their meals with just one photo.

What it does NutriScan helps users get quick and accurate nutrition information from a single image. Here’s what happens behind the scenes: The user opens the NutriScan mobile app and takes a photo of their food.

The app sends this image to a backend running on Google Cloud Run.

The backend activates the NutriScan Sequential Agent, which is built using Google’s ADK.

Inside that agent, two smart sub-agents work together:

Vision Agent uses the Gemini 2.5 Flash model to identify the food items in the image and understand what’s on the plate.

Nutrition Analysis Agent takes that output and creates a simple text summary that includes calories, protein, carbs, and fat.

The system then sends the result back to the mobile app, where the user can read it instantly. If the photo cannot be processed, the app shows a friendly message explaining what went wrong. How it works: Frontend – Mobile App Built using React Native and Expo.

Allows the user to capture a photo or pick one from their gallery.

Uses Axios to send the image securely to the backend as multipart form data.

Displays a smooth loading animation while processing.

Shows results in a clean, modern card layout.

Runs on Android through Expo Go for quick development and testing.

Backend – Cloud Run Service Deployed on Google Cloud Run as a Docker container for scalability and reliability.

When it receives an image, it calls the NutriScan Agent workflow.

NutriScan Agentic Workflow – Google ADK Built using Google’s Agent Development Kit (ADK) to create a modular, multi-agent workflow.

The Sequential Agent acts as the main controller that runs two Gemini 2.5 Flash agents in order:

Vision Agent: Understands the image, identifies the food items, and estimates portion sizes.

Nutrition Analysis Agent: Converts the detected items into nutritional facts.

Once the analysis is done, the agent returns a simple text message containing all the relevant nutrition details.

The backend formats this message into a JSON response and sends it back to the app.

Tech Stack

Layer Tools and Technologies Frontend React Native, Expo, Axios, Expo Image Picker Backend FastAPI (built by adk deploy) AI Agents Google ADK, Gemini 2.5 Flash Deployment Docker, Google Cloud Run Development Tools Node.js, npm, VS Code

Challenges Learning Google Cloud Run and Google ADK from scratch. As the technology was new, the LLMs like ChatGPT and Gemini were struggling to help with the coding. YouTube, documentation, and forums were our best friends in this journey.

Getting the app and backend to communicate smoothly over the same network.

Handling image uploads from React Native.

Learning how to build a structured agent workflow using Google ADK for the first time.

Balancing speed and reliability in the Cloud Run environment.

Accomplishments Created a working end-to-end system where a user can take a photo and instantly get nutrition details.

Designed a fully functional Sequential Agent using the Google ADK.

Successfully deployed our containerized backend on Google Cloud Run.

Built a clean, simple mobile interface that makes advanced AI feel friendly and accessible.

Future work: Connect the Nutrition Analysis Agent to real-world nutrition databases such as USDA FoodData Central.

Add support for multi-dish detection and serving-size adjustments.

Integrate voice interaction to make the experience even more accessible.

Deploy the agentic workflow to Vertex AI Agent Builder for enterprise-scale performance.

Build a user dashboard that tracks meal history and daily nutrition summaries.

Built With

Share this project:

Updates