Guido_tourist_guide

"Meet your personal AI Travel Guide: simply tell me which city you’re heading to, and I’ll generate a comprehensive travel guide for you.

Comment

Guido — AI Cognitive Travel Concierge

Inspiration

Travel planning is often a logistical burden that detracts from the magic of discovery.

We envisioned a “Cognitive Concierge” — an agentic system that doesn't just list destinations but visually stages them.

Our goal was to create an immersive pre-trip experience where AI acts as a digital scout:

  • Finding hidden gems
  • Illustrating them in real time
  • Bridging the gap between imagination and departure

Instead of browsing dozens of travel websites, users receive a curated digital travel magazine generated instantly by AI.

What It Does

Guido is an AI-powered travel consultant that transforms natural language queries into a high-fidelity interactive digital magazine.

It can:

  • Research authentic travel locations using live web data
  • Generate cinematic photography matching the user's aesthetic
  • Drive a dynamic React interface
  • Turn a static chat into an Agentic UI

The interface evolves dynamically based on what the AI agents discover.

How We Built It

Multi-Agent Architecture

We implemented a multi-agent orchestration system using:

  • Google Agent Development Kit (ADK)
  • CopilotKit

Intelligence Layer

Orchestrated via:

  • SequentialAgent
  • ParallelAgent

The Agents

The Researcher

A guide_agent using Google Search Grounding to ensure:

  • factual information
  • up-to-date local insights
  • real-world locations

The Visualizers

Parallel agents using:

gemini-3.1-pro-image

They generate:

  • high-resolution photography
  • architectural imagery
  • lifestyle visuals
  • consistent cinematic aesthetics

The Editor

A guide_redactor agent that:

  • aggregates all outputs
  • synthesizes multimodal results
  • outputs structured data in a strict JSON schema

Frontend

The UI is built using:

  • Next.js
  • React

It implements the A2UI protocol (Agent-to-UI) v0.9, allowing the AI agent to:

  • control frontend components
  • dynamically modify the interface
  • transform responses into structured UI elements

This enables the shift from chat interface → agent-driven UI.

DevOps & Infrastructure

Dependency Management

  • uv for high-speed Python dependency resolution

Deployment

  • Google Cloud Run

Asset Storage

  • Google Cloud Storage (GCS)

Used for:

  • hosting AI-generated images
  • persistent asset delivery

Challenges We Ran Into

Protocol Synchronization

Aligning schemas between:

  • Python Pydantic
  • React Zod

during the migration to A2UI v0.9.

Even small mismatches caused UI crashes due to invalid streaming data.

Model Versioning Latency

We encountered 404 NOT_FOUND errors due to the rapid evolution of the Gemini 3 API.

Solution:

  • built a dynamic model-mapping layer
  • automatically resolves v1beta → v1 endpoint transitions

Artifact Visibility

AI-generated images initially streamed as base64 strings, which caused instability.

Solution:

  • implemented custom middleware
  • registered images as ADK Artifacts
  • served via authenticated persistent URLs

Accomplishments We're Proud Of

Latency Optimization

We parallelized:

  • research
  • visual generation

This drastically improved:

Time to First UI Render

The system feels responsive despite heavy AI workloads.

Seamless State Management

Successfully propagated:

  • API keys
  • service accounts

through a secure .env pipeline to Google Cloud production infrastructure.

Visual Coherence

We engineered a shared style constraint injected into every visual agent.

Result:

  • consistent editorial look
  • cinematic visual identity
  • magazine-style image cohesion

What We Learned

The Power of Agentic UI

AI is evolving beyond chat interfaces.

Agents that control UI components directly provide:

  • better usability
  • clearer results
  • richer experiences

Infrastructure Reliability

We learned that the plumbing matters as much as the AI.

Switching to:

  • uv
  • GCS

significantly improved stability.

Grounding Is Non-Negotiable

For travel applications:

ungrounded generation is dangerous.

Google Search Grounding ensures:

  • real locations
  • current availability
  • factual data

Without it, hallucinations would break the experience.

What's Next for Guido

Native Video Integration

Integrating Google Veo to generate:

  • cinematic 4K travel clips
  • immersive destination previews

MCP Expansion

Implementing Model Context Protocol (MCP) to access:

  • user calendar
  • past travel history
  • preferences

This enables hyper-personalized itineraries.

Voice-First Mode

Using Gemini Live Mode to enable:

  • real-time voice conversation
  • dynamic UI updates while speaking

Users will plan trips hands-free while watching the UI evolve visually.

Built With

Share this project:

Updates

Submission history