NomadAgent

Describe your trip in plain language, watch AI research it in real time, and receive a verified itinerary.

Features Tech Stack Architecture

Overview

NomadAgent is an AI-powered travel research agent that turns natural-language trip descriptions into fully verified itineraries. You describe where you want to go, what you're interested in, and how long you're staying — then watch in real time as the agent plans research tasks, searches the web, extracts structured venue data, and compiles a day-by-day itinerary with coordinates, opening hours, and source citations.

Built with a Flutter mobile frontend and a Python FastAPI backend powered by LangGraph and Google Gemini, NomadAgent demonstrates how agentic AI pipelines can solve real-world planning problems end to end.

Features

• Natural Language Trip Input — Describe your trip in plain English; the agent parses destination, duration, and interests automatically
• Multi-Step Agent Pipeline — LangGraph orchestrates a planner → researcher → extractor → compiler workflow with full state management
• Real-Time Streaming — Server-Sent Events (SSE) stream thought logs, venue verifications, and self-corrections to the app as they happen
• Intelligent Web Research — Tavily search with hybrid relevance scoring filters irrelevant results using confidence + destination keyword matching
• LLM-Powered Data Extraction — Gemini extracts structured venue data (coordinates, addresses, hours, descriptions) from raw web content
• Venue Verification — Tiered verification scoring with type-specific weight tables so nature venues aren't penalized for lacking opening hours
• Interactive Map View — Flutter Map renders all venues with accurate coordinates on an interactive OpenStreetMap layer
• Day-by-Day Itinerary — Compiled itinerary groups venues by day with travel logistics and time estimates
• PDF Export & Sharing — Generate and share polished PDF itineraries directly from the app
• Parallel Extraction — asyncio.gather runs concurrent LLM calls across venue tasks, cutting extraction time by ~60%
• Bounded Event History — Monotonic cursor + rolling buffer keeps streaming memory-bounded during long generations
• Graceful Error Handling — Failed search tasks and 503 errors are skipped without blocking the pipeline

Demo

Home Screen	Home with Prompt	Event Streaming	Itinerary View	Map View

Video Demo: Full end-to-end generation for this test prompt:

"I'm planning a weekend trip to Barcelona next month. I'm interested in food and history. I have 2 days and a budget of $500."

Generation time: ~46 seconds (edited for brevity)

Watch the demo video

Tech Stack

Architecture

┌─────────────────────────────────────────────────────────┐
│                    Flutter Mobile App                    │
│         (Riverpod · GoRouter · Flutter Map)              │
│                                                         │
│  Home → Generation (SSE) → Itinerary → Map → PDF Export │
└────────────────────────┬────────────────────────────────┘
│ SSE (Server-Sent Events)
▼
┌─────────────────────────────────────────────────────────┐
│                   FastAPI Backend                        │
│              POST /api/v1/generate                       │
│              GET  /api/v1/health                         │
├─────────────────────────────────────────────────────────┤
│                 LangGraph Agent Pipeline                 │
│                                                         │
│  ┌──────────┐   ┌────────────┐   ┌───────────┐         │
│  │ Planner  │──▶│ Researcher │──▶│ Extractor │         │
│  │ (Gemini) │   │  (Tavily)  │   │  (Gemini) │         │
│  └──────────┘   └────────────┘   └─────┬─────┘         │
│                                        │               │
│                                  ┌─────▼─────┐         │
│                                  │  Compiler  │         │
│                                  │  (Gemini)  │         │
│                                  └───────────┘         │
└─────────────────────────────────────────────────────────┘

Pipeline flow:

1. Planner — Parses the user prompt, identifies the destination, duration, and interest categories, then generates up to 3 focused research tasks
2. Researcher — Executes Tavily web searches for each task with relevance scoring and result deduplication
3. Extractor — Runs parallel Gemini calls to extract structured venue data (name, coordinates, address, hours, description, source URL) from raw search content
4. Compiler — Assembles verified venues into a day-by-day itinerary with logistics and time estimates