Céleste - Multi-Agent AI Shopping Assistant

A sophisticated multi-agent AI system built for the GKE Turns 10 Hackathon. Céleste demonstrates advanced agentic AI coordination using three specialized agents that collaborate to provide intelligent shopping assistance through natural language conversations.

Live Demo: http://34.54.94.175/home

Multi-Agent Architecture

Céleste implements a true agentic AI system where autonomous agents coordinate to solve complex shopping queries:

Agent Orchestrator

Central coordinator that manages agent communication and synthesizes responses from multiple AI agents working in parallel.

Search Agent

• Analyzes customer queries using Google Gemini AI
• Performs intelligent product matching against catalog data
• Classifies intent (product search, style advice, price inquiry, etc.)
• Returns ranked product results with relevance scoring

Inventory Agent

• Simulates real-time stock level checking
• Analyzes demand patterns and trending items
• Provides inventory recommendations and alerts
• Tracks seasonal availability patterns

Recommendation Agent

• Generates personalized suggestions based on user context
• Maintains conversation history for improved recommendations
• Creates contextual follow-up actions
• Calculates personalization scores for enhanced user experience

Technology Stack

• Language: Go 1.25+ with Gorilla Mux routing
• AI Engine: Google Gemini 2.5 Flash for natural language processing
• Agent Framework: Custom implementation inspired by Agent Development Kit (ADK) patterns
• Message Protocol: MCP-inspired inter-agent communication
• Deployment: Google Kubernetes Engine (GKE) Autopilot
• Container Registry: Google Artifact Registry
• Networking: Kubernetes Ingress with external load balancer

Features

Conversational AI Interface

• Natural language product queries with context understanding
• Real-time agent coordination visualization
• Mobile-responsive progressive web interface
• Animated agent activity indicators showing multi-agent workflow

Intelligent Product Discovery

• Semantic search across product catalogs
• Intent classification and query analysis
• Synonym matching and category-based filtering
• Personalized recommendations based on conversation history

Real-time Agent Coordination

• Autonomous agent decision-making and task distribution
• Inter-agent message passing and state synchronization
• Workflow visualization showing agent participation
• Graceful error handling and agent fallback mechanisms

API Endpoints

GET /

Welcome message and system status

GET /home

Responsive web interface for interactive agent demonstrations

• Demo URL: http://34.54.94.175/home

GET /health

System health monitoring endpoint that returns agent status and availability

• Lists all registered agents (search_agent, inventory_agent, recommendation_agent)
• Reports system health status
• Shows active agent count for monitoring
• Example: http://34.54.94.175/health

GET /agents

Agent registry endpoint that lists all available agents and their capabilities

• Returns live agent inventory with names and status
• Shows total agent count
• Useful for debugging and system monitoring
• Example: http://34.54.94.175/agents

POST /chat

Multi-agent query processing endpoint that triggers the full agent workflow

{
  "query": "I need winter boots for hiking",
  "user_id": "demo_user"
}