CartSense AI

CartSense AI is a proof-of-concept for the Online Boutique that detects abandoned fashion carts, uses Gemini for reasoning, and generates mock recommendations, all on GKE with MCP and agents.

Comment

📖 About the Project

🌟 Inspiration

Our team was inspired by the challenge prompt: “Give Microservices an AI Upgrade.” We wanted to explore how an existing microservice-based application—Online Boutique—could be enhanced without altering its core code.

E-commerce is full of scenarios where customers add items to their cart but never complete the purchase. This provided us with a relatable, high-impact use case: abandoned carts. We decided to focus specifically on fashion items because they often involve style, preference, and recommendation opportunities, making them a perfect playground for AI-driven personalization.

🛠️ How We Built It

Our solution, CartSense AI, is built as a proof-of-concept extension that integrates with the Online Boutique demo application using external, containerized agents.

Architecture

  • Google Kubernetes Engine (GKE): All agents and services are deployed as containers in GKE, running alongside the Online Boutique.
  • Model Context Protocol (MCP): We use MCP to interact with the Boutique’s APIs, particularly for monitoring cart states and product IDs.
  • Agent Orchestration:
    • Cart Monitor Agent: Periodically or event-driven, detects when carts contain fashion items.
    • AI Reasoning Filter: Determines if a cart has been abandoned.
    • Recommendation Agent: Uses Gemini models to suggest complementary products based on cart contents.
    • Notification Generator: Creates mock push notifications/emails with the recommendations (since this is demo data, no actual users are contacted).
  • Google Gemini Models: Provide reasoning and contextual personalization to ensure the recommendations feel intelligent rather than static.

📚 What We Learned

  • Agentic AI is modular and powerful: We realized how external agents can bring intelligence into existing systems without requiring invasive changes.
  • MCP unlocks flexibility: Using MCP to bridge APIs allowed us to treat the Online Boutique as a black box while still interacting with it meaningfully.
  • Kubernetes orchestration: We sharpened our skills in deploying and managing multiple agent containers on GKE.
  • AI reasoning ≠ traditional rules: Adding Gemini to filter abandoned carts and generate recommendations showed us the difference between simple if/else rules and context-aware reasoning.

⚔️ Challenges We Faced

  • No real users: Since Online Boutique is a demo app, we worked only with sample carts and synthetic data, so our notifications and recommendations are mock outputs.
  • API limitations: The Boutique APIs aren’t designed for real-world AI workflows, so we had to carefully design agents to work within those boundaries.
  • Agent orchestration: Coordinating multiple agents (monitoring, reasoning, recommending, notifying) and ensuring they communicated correctly required careful workflow design.
  • Deployment on GKE: Managing containers, services, and ensuring our setup scaled correctly took extra debugging and iteration.

🚀 Takeaway

CartSense AI demonstrates how agent-driven intelligence can be layered onto microservice applications. Even in a fictional boutique with sample data, the concept illustrates how future e-commerce systems could:

  • Detect user behaviors (like abandoned carts).
  • Reason about intent and context.
  • Deliver personalized, AI-powered suggestions.

In short, we learned that agents + microservices + AI = a powerful recipe for innovation.

Built With

Share this project:

Updates