ערכת נושא
Architecture Document · v2.0 · 2026
t.Co.
Multi-Agent Architecture for Hyper-Personalized Travel
STAGE 1 Before Trip AGT-01 · Preference Profiling AGT-02 · Itinerary Creation AGT-03 · Booking & Mgmt STAGE 2 During Trip AGT-04 · Location Context AGT-05 · Schedule Optimizer AGT-06 · Language & Culture STAGE 4 Long-Term AGT-09/10/11 STAGE 3 Post Trip AGT-07/08 CORE Personal AI Orchestrator
11 Agents · 4 Stages · 1 Evolving Companion
Table of Contents
  1. Executive Summary
  2. System Architecture
  3. Agent Catalog
  4. Core Orchestrator
  5. Technical Stack
  6. Demo Scope
  7. Exhibition Experience

01 — Executive Summary

A Lifelong Companion for Every Journey

The Personal AI Travel Concierge is a multi-agent system designed to deliver hyper-personalized travel assistance at every stage of the journey — from initial inspiration to long-term travel memory. Unlike conventional travel apps that treat each booking as an isolated transaction, this system builds a continuous, evolving model of the traveler's preferences, behaviors, and aspirations.

At its core, an orchestrating AI coordinates eleven specialized agents across four distinct stages: pre-trip planning and booking, real-time in-trip guidance, post-trip curation and reflection, and long-term companion intelligence. Each stage feeds the next, creating a feedback loop that makes the system smarter and more attuned with every journey.

This architecture document provides a complete specification of the system's agent topology, data flows, technical dependencies, and exhibition implementation scope.

02 — System Architecture

The Four-Stage Agent Ecosystem

The system is organized around a central orchestrating intelligence — the Core Personal AI — which coordinates four specialized stage clusters. Data and context flow continuously between stages, with a feedback loop from Stage 4 back to Stage 1.

STAGE 1 Before Trip AGT-01 Preference Profiling AGT-02 Itinerary Creation AGT-03 Booking & Management LIVE ×2 STAGE 2 During Trip AGT-04 Location Context AGT-05 Schedule Optimizer AGT-06 Language & Culture LIVE ×1 STAGE 3 Post Trip AGT-07 Memory Curation AGT-08 Feedback & Learning STAGE 4 Long-Term Value AGT-09 Loyalty Optimization AGT-10 Wellness Monitoring AGT-11 Trip Anticipation CORE Personal AI Orchestrator plan → execute insights → profile

03 — Agent Catalog

Eleven Specialized Agents

S1

Before Trip — Planning & Booking

Transforms raw travel intent into a confirmed, personalized itinerary with all reservations secured.

AGT-01
Preference Profiling Agent
Builds a rich, structured model of the traveler's style, budget thresholds, interests, and constraints through natural conversation and behavioral inference.
Inputs
  • User conversations & stated preferences
  • Past trip history
  • Budget & date constraints
Outputs
  • Structured user profile (JSON)
  • Travel style taxonomy
  • Interest & constraint vectors
GPT-4oVector DBClaude
✓ Implemented
AGT-02
Itinerary Creation Agent
Generates complete day-by-day travel plans optimized for the user's profile — balancing must-sees with hidden gems, respecting energy levels and budget.
Inputs
  • User profile (AGT-01 output)
  • Destination, dates, group size
  • Budget envelope
Outputs
  • Structured itinerary (JSON + readable)
  • Activity cards with location & cost
  • Map-ready coordinate set
Google PlacesOpenAIFoursquare
✓ Implemented
AGT-03
Booking & Management Agent
Executes reservations across flights, hotels, and restaurants — organizes all confirmations, tickets, and calendar entries into a single travel pack.
Inputs
  • Confirmed itinerary (AGT-02)
  • Payment preferences & loyalty accounts
Outputs
  • Confirmed bookings & PNR codes
  • Calendar events & reminders
  • Document pack (tickets, vouchers)
AmadeusBooking.comOpenTable
○ Conceptual
S2

During Trip — Real-Time Guidance

Acts as an always-on local guide — sensing context, resolving disruptions, and bridging language barriers in real time.

AGT-04
Location Context Agent
Monitors real-time position, crowd levels, and time-of-day to surface proactive, hyper-relevant suggestions before the traveler even asks.
Inputs
  • GPS coordinates & heading
  • Timestamp & local conditions
  • Crowd density signals
Outputs
  • Ranked nearby suggestions
  • Contextual alerts & detour advice
Google MapsHERE Maps
○ Conceptual
AGT-05
Schedule Optimization Agent
Continuously monitors for disruptions — delays, weather, closures — and automatically re-plans, triggering rebooking actions where needed.
Inputs
  • Live itinerary state
  • Flight status feeds
  • Weather & event alerts
Outputs
  • Updated itinerary
  • Rebooking actions via AGT-03
  • Proactive notifications
FlightRadarOpenWeather
○ Conceptual
AGT-06
Language & Culture Agent
Provides real-time spoken translation and culturally nuanced etiquette guidance — from pronunciation help to awareness of local customs.
Inputs
  • User speech or typed text
  • Current GPS location
  • Target language context
Outputs
  • Spoken & written translations
  • Cultural etiquette tips
  • Pronunciation audio
Whisper STTDeepLElevenLabs
✓ Implemented
S3

Post Trip — Memories & Insights

Transforms raw trip data into curated memories and actionable insights that feed the next journey.

AGT-07
Memory Curation Agent
Organizes photos, visited locations, and experiences into a coherent travel narrative — ready to share or preserve as a personal travel journal.
Inputs
  • Trip photos & media
  • Location history & timestamps
  • Booking records
Outputs
  • Organized photo albums
  • Narrative travel journal
  • Social media–ready posts
Google PhotosGPT-4 Vision
○ Conceptual
AGT-08
Feedback & Learning Agent
Captures explicit ratings and implicit behavioral signals to refine the user's travel profile — ensuring every future trip is better calibrated.
Inputs
  • Star ratings & comments
  • Behavioral signals (time spent, skips)
  • Post-trip survey responses
Outputs
  • Updated preference profile
  • Insights report per destination
  • Recommendation model refinement
Sentiment AnalysisVector DB
○ Conceptual
S4

Long-Term Value — The Evolving Companion

Grows alongside the traveler over years — optimizing loyalty, safeguarding wellness, and anticipating the next adventure.

AGT-09
Loyalty Optimization Agent
Tracks points, miles, and status across all programs — maximizing reward redemptions and steering booking decisions toward highest long-term value.
Inputs
  • Booking history
  • Loyalty program memberships
  • Upcoming trip parameters
Outputs
  • Redemption recommendations
  • Program upgrade suggestions
  • Optimal booking timing alerts
Airline APIsHotel Loyalty
○ Conceptual
AGT-10
Wellness Agent
Monitors trip intensity, jet lag, and physical activity — proactively suggesting rest or recovery to maintain optimal well-being throughout the journey.
Inputs
  • Activity & step data
  • Trip schedule density
  • Health preferences & conditions
Outputs
  • Wellness alerts & rest nudges
  • Schedule lightening suggestions
HealthKitWearable APIs
○ Conceptual
AGT-11
Future Trip Anticipation Agent
Proactively surfaces destination proposals — drawing on travel history, seasonal opportunities, trending destinations, and the traveler's evolving interests.
Inputs
  • Full travel history & feedback
  • Trending destination signals
  • Seasonal & pricing calendars
Outputs
  • Ranked destination proposals
  • Optimal travel timing windows
  • Estimated budget previews
Web SearchGPT-4o
○ Conceptual

04 — Core Orchestrator

The Intelligence at the Center

The Core Personal AI is not a passive router — it is the persistent intelligence that gives the system its character. It maintains a unified, evolving model of the traveler, mediates between agents, and ensures every interaction feels cohesive and intentional.

Core Responsibilities

  • Maintains the unified traveler profile across all sessions and trips
  • Routes user requests to the appropriate specialized agent
  • Manages conversation context window across all stages
  • Resolves conflicts between competing agent recommendations
  • Synthesizes multi-agent outputs into a single coherent voice
  • Triggers proactive outreach when context warrants it

Technology Foundation

  • LLM backbone: Anthropic Claude / OpenAI GPT-4o
  • Multi-agent coordination: LangGraph / CrewAI
  • Persistent memory: PostgreSQL + Pinecone vector store
  • Real-time event bus: WebSocket + message queue
  • Prompt engineering: structured role cards per agent
  • Safety layer: content moderation + PII filtering

05 — Technical Stack

Infrastructure & Integrations

AI / LLM Layer
GPT-4oClaude 3.5LangGraphCrewAI
Voice Interface
OpenAI WhisperElevenLabs TTSWebSpeech API
Location Services
Google MapsFoursquareHERE Maps
Travel & Booking
Amadeus APIBooking.comOpenTable
Data & Memory
PostgreSQLPineconeRedisS3
Frontend
React 18Tailwind CSSFramer MotionMapbox GL
Backend
Node.jsPython FastAPIWebSocketsDocker
External Intelligence
DeepL TranslateOpenWeatherFlightAware

06 — Demo Scope

Exhibition Implementation vs. Full Vision

The exhibition demo presents a working proof-of-concept for the three highest-impact agents, demonstrating the core interaction loop — from preference intake through personalized itinerary generation — with voice interaction and live visual output.

✓ Implemented for Exhibition
  • AGT-01 — Preference Profiling (conversational onboarding)
  • AGT-02 — Itinerary Creation (real-time generation + map viz)
  • AGT-06 — Language & Culture Agent (live translation)
  • Voice interaction — speech-to-text via Whisper
  • AI voice output — spoken responses via ElevenLabs
  • Real-time itinerary visualization on large screen
  • Interactive agent map — architecture visible as UI layer
○ Conceptual / Future Work
  • AGT-03 — Actual booking execution (payment integration)
  • AGT-04 — Live GPS recommendations
  • AGT-05 — Real disruption handling & rebooking
  • AGT-07 — Media curation & social sharing
  • AGT-08 — Multi-trip feedback learning loop
  • AGT-09 — Cross-program loyalty optimization
  • AGT-10 — Wearable wellness integration
  • AGT-11 — Long-horizon trip prediction engine

07 — Exhibition Experience

The Installation: Travel Intelligence, Live

The exhibition presents the system as an immersive, large-format installation. A visitor approaches a floor-standing screen — the AI companion initiates a natural conversation about their dream destination. Within minutes, a personalized itinerary materializes in real time.

01

Approach & Engage

Visitor approaches the screen. The AI avatar activates and greets them. Keyboard available as fallback.

02

Dream Destination

The AI asks: "Where have you always wanted to go?" Preference profiling begins in natural dialogue.

03

Live Itinerary

A map materializes. Day-by-day activities appear in real time as the Itinerary Agent thinks.

04

Language Demo

The Language Agent demonstrates pronunciation help and cultural context for the destination.

05

Architecture Reveal

The background agent map animates — showing which agents activated and how decisions were made.

t.Co. · Architecture Document v2.0 · M.Design Thesis · HIT 2026

11 Agents · 4 Stages · 1 Evolving Companion