How a 5-Day AI Agents Intensive Course Helped Me Build FoodMate, A Multi Agent Food Concierge

Source: Dev.to

Introduction

When I signed up for the Google × Kaggle 5‑Day AI Agents Intensive, I didn’t expect to finish a polished project in a single week. The course structure, hands‑on codelabs, and community support gave me the confidence to build something simple, useful, and real: FoodMate, a multi‑agent food recommendation assistant.

The biggest idea that stuck with me was that agents are teams, not single models. Separating tasks into agents for preference, nutrition, and verification made the system easier to reason about, test, and extend. The course showed practical patterns:

• Tool use & grounding – how to fetch external info safely
• Function calling and structured outputs – so downstream apps can use results
• Agent orchestration – agents communicating in sequence and in parallel

Why FoodMate

Food is personal. In Sri Lanka, small constraints like time, budget, and local ingredients shape daily choices. I built FoodMate to:

• Propose quick, culturally relevant meal ideas
• Check a simple nutrition heuristic
• Provide healthier swaps when appropriate

It’s intentionally tiny so it runs even with limited resources.

Architecture

• Preference Agent – captures user tastes and constraints
• Recipe Agent – retrieves suitable recipes from a local database
• Nutrition Agent – evaluates the nutritional profile of suggested meals
• Aggregator – combines outputs and presents the final recommendation

Tech stack: Kaggle Notebook, Python, Gemini / ADK patterns (with a fallback local recipe DB for quota issues).

Demo

• Notebook: https://www.kaggle.com/code/shiyazamal/capstone-foodmate-multi-agent-food-recommendati
• Video walkthrough: https://youtu.be/-3B8amyQBKc

What Surprised Me

Working through the codelabs made me realize agents don’t need to be exotic. Even small, reliable modules that do clear jobs create a system that feels “intelligent” and actually helpful. The community support in the Kaggle Discord was huge—people helped debug codelabs and suggested better prompt designs.

Next Steps

• Add user memory to remember preferences over time
• Deploy the agent to Agent Engine or Cloud Run
• Connect to local restaurant menus for live suggestions
• Build a simple mobile UI so people can use FoodMate while shopping

This intensive didn’t just teach me tools; it changed my mindset. I moved from “how do I query a model?” to “how do I design a small team of agents that reliably solves one useful problem?” That’s what I’ll be building next.

Links

• Notebook: https://www.kaggle.com/code/shiyazamal/capstone-foodmate-multi-agent-food-recommendati
• Demo video: https://youtu.be/-3B8amyQBKc

If you’d like feedback on your project or want to collaborate, ping me on Kaggle: shiyazamal.