AI Trip Planner - NomadAI

– NOMADAI: AI Trip Planner Agent –

– Overview –
AI Trip Planner Agent is a fully cloud-native, agentic travel planning application that leverages a self-hosted open-source LLM (Gemma 3 1B) deployed on Google Cloud Run with GPU acceleration. The system provides users with comprehensive, real-time travel itineraries, integrating live data from multiple APIs (weather, places, currency, expenses) and offering a modern, user-friendly web interface.
————————————————–

– Core Requirements –

1. LLM Deployment

Model: Google Gemma 3 1B (open-source, lightweight, cost-efficient)
Serving Framework: Ollama
Deployment: Google Cloud Run with NVIDIA L4 GPU, Europe region, non-zonal redundancy, single instance for cost control
API Endpoint: Exposed securely for agent inference, using –allow-unauthenticated for hackathon demo access

2. Agent Development

Framework: LangGraph (LangChain ecosystem)
Workflow: React-style agentic workflow (Reason + Action), with multi-tool orchestration
Tools Integrated:

• Weather (OpenWeatherMap)
• Place search (Google Places, Foursquare, Tavily fallback)
• Currency conversion (Exchange Rate API)
• Expense calculation
  Custom LLM Integration: Agent is adapted to interact with the self-hosted Gemma 3 1B model via HTTP, with custom tool-calling logic to compensate for lack of native function-calling in open models

3. Agent Deployment

Backend: FastAPI, deployed on Cloud Run (serverless, scalable, secure)
Frontend: Streamlit, deployed on Cloud Run for a beautiful, interactive user experience
API Key Management: All secrets are managed via Cloud Run environment variables (never committed to git)
End-to-End Cloud Native: All components (LLM, agent, frontend) are deployed as independent Cloud Run services
————————————————–

– Unique Features & Innovation –
Self-Hosted LLM: No reliance on proprietary APIs; full control, privacy, and cost efficiency
Agentic Reasoning: Multi-step, multi-tool planning for rich, actionable itineraries
Real-Time Data: Integrates live weather, places, and currency data for up-to-date recommendations
Modern UI: Streamlit frontend with query templates, live status, and downloadable plans
Fallback Resilience: Multiple data sources (Google, Foursquare, Tavily) ensure robust results
User Experience: One-click templates, markdown export, and clear guidance for best results
————————————————–

– Tech Stack –
Frontend: Streamlit (Python)
Backend: FastAPI (Python)
Agent Framework: LangGraph (LangChain)
LLM Serving: Ollama
Cloud Platform: Google Cloud Run (GPU for LLM, CPU for agent/frontend)
APIs: OpenWeatherMap, Google Places, Foursquare, Exchange Rate API, Tavily
————————————————–

– Business Value –
Personalized, actionable travel plans in seconds
Cost-effective (open model, minimal GPU usage, serverless scaling)
Privacy and control (no user data leaves your cloud)
Extensible (easy to add new tools or swap models)
Modern, delightful user experience (beautiful UI, instant feedback, downloadable plans)
————————————————–

– Performance Optimization –
Model Selection: Gemma 3 1B for fast inference and low GPU cost
Cloud Run Tuning: Single GPU, non-zonal redundancy, concurrency settings for optimal scaling
API Caching: (Optional) Can be added for repeated queries
Frontend/Backend Separation: Ensures responsive UI even under load
————————————————–

– Security Best Practices –
No secrets in git: All API keys managed via environment variables
Cloud Run isolation: Each service runs in its own secure container
No public LLM endpoints: Only agent backend calls the LLM
Input validation: All user input is sanitized and validated
————————————————–

– Error Handling –
Graceful fallback: If a primary API fails, the agent tries alternatives
User feedback: Clear error messages and guidance in the UI
Logging: All errors are logged for debugging and monitoring
————————————————–

– Summary –
AI Trip Planner Agent is a robust, innovative, and fully cloud-native solution that demonstrates best practices in LLM deployment, agentic reasoning, and user-centric design. It is ready for real-world use and further extension!
————————————————–