Nora Labs
Autonomous Agentic Tourism & Multi-Agent Planning Engine
Tourthon 2026 AI-driven multi-agent platform for bespoke Sri Lankan travel itineraries, live driver dispatch, and instant dynamic booking.
Overview
Nora Labs is an enterprise-grade agentic travel platform engineered for Tourthon 2026. It revolutionizes travel planning in Sri Lanka by replacing static tour packages with autonomous collaborative AI agents. The multi-agent backend dynamically queries real-time regional weather, curated cultural attractions, verified private drivers, and luxury accommodation to formulate hyper-personalized, conflict-free travel itineraries.
Design decisions
The choices that shaped the system, and why.
- 1
Why Google ADK over LangChain/LangGraph?
LangChain's rapidly shifting abstractions introduce breaking dependency changes and overhead. Google ADK provides lightweight, deterministic state-machine loops tailored for Gemini 1.5 Flash function calling, delivering sub-200ms agent decision turns.
- 2
Why Gemini 1.5 Flash for the reasoning loop?
Tourism route planning requires processing multi-day logistical constraints across dozens of cultural sites simultaneously. Gemini 1.5's massive context window and fast token generation eliminate the need for complicated RAG chunking while reducing LLM inference costs by 80%.
- 3
Decoupled Agent Architecture
Split into specialized agents (Trip Architect, Driver Dispatcher, Budget Controller, Itinerary Validator) rather than a monolithic prompt, preventing hallucinated travel times between distant Sri Lankan cities (e.g. Colombo to Ella).
- 4
Dual-Portal Frontend Separation
Built a traveler-facing dynamic wizard and an isolated administrative driver-dispatch dashboard to prevent state pollution and enforce role-based access control.
Architecture
Multi-Agent Orchestration Engine
Built with Google ADK and FastAPI, featuring specialized autonomous agents: Trip Architect Agent, Driver Matching Agent, Itinerary Optimizer Agent, and Budget Controller.
Gemini Reasoning Loop
Employs structured Chain-of-Thought prompting and function calling to evaluate multi-city constraints (e.g., Sigiriya -> Kandy -> Ella route efficiency).
Dual-Frontend Ecosystem
Features a modern traveler portal (React/Vite/Tailwind) and an administrative driver-dispatch dashboard.
Microservices Infrastructure
Containerized with Docker Compose, providing resilient REST endpoints, Redis caching, and automated health checks.
DevOps & deployment
Packaged with multi-stage Dockerfiles. Configured for zero-downtime deployment on Google Cloud Run with Cloud Memorystore Redis, or as scalable ECS Fargate tasks behind an AWS Application Load Balancer.
How it ships · Google Cloud Run (us-central1) · Vertex AI · GCS-mounted SQLite
- 1git push
- 2Cloud Build: docker build (tag :SHORT_SHA + :latest)
- 3Push to Artifact Registry
- 4gcloud run deploy --max-instances=1
- 5Vertex AI credentials via runtime service account
Features
- Autonomous Natural Language Trip Wizard generating day-by-day itineraries with travel times and scenic stops.
- Dynamic verified driver assignment with vehicle classification (Tuk-tuk, Sedan, Van, 4x4) and language matching.
- Interactive route mapping with waypoint optimization across cultural triangle, hill country, and southern coast.
- Integrated PDF & DOCX proposal and itinerary exporter with cost breakdowns and booking vouchers.
Screenshots
Related writing
DevOps · 4 min
Cloud Run Problems I Actually Hit — and the Morning My Demo Was Gone
Missing service accounts, region-locked models, disks that vanish, one-writer SQLite, schedulers that stop — and the build-time check that found Nora Labs returning Google's 404.
29 Oct 2026
Deep dive · 4 min
Tourism 10: One File the Owner Can Edit
Resorts, guides and wellness centres change prices and phone numbers — not code. Put every editable fact in one typed data file.
20 Oct 2026
Deep dive · 4 min
Tourism 09: An AI Trip Planner That Still Works When the AI Doesn't
Specialist agents for search, itinerary, pricing and dispatch — each calling the same Python functions as a deterministic fallback.
19 Oct 2026
Deep dive · 4 min
Tourism 08: Transparent Pricing and OTA Deep Links
A deterministic nightly rate from season, occupancy and lead time — with its reasoning attached — and honest Agoda / Booking.com search links instead of scraping.
18 Oct 2026