Gipsyy Ecommerce

Gipsyy sells intercity bus travel across Portugal, Spain and Italy, three markets with different languages, tax rules and payment habits, on top of a checkout that has to price and lock seats in real time. The brief was to rebuild the booking platform so it could scale to new countries without a rewrite each time: multi-leg itineraries, per-market payments, and a checkout fast enough that travelers don't abandon it, all while a small team kept shipping on web and mobile at once.

I split the system by domain instead of by surface. A polyglot service layer (.NET 7 for the booking, inventory and pricing core, Ruby on Rails for supporting domains) sits behind one API contract, consumed by a Next.js web app and a React Native mobile app that share TypeScript types. Payments and confirmation are event-driven, so a slow payment provider never blocks the seat-locking path. On top sits an AI layer, RAG retrieval and LangGraph agents, that automates multilingual support and generates itinerary content.

• Domain-split microservices behind a gateway: the booking hot path (search, seat inventory, pricing) scales independently from the slower supporting services.
• Confirmation runs on events through RabbitMQ, so Adyen settlement is decoupled from seat-locking. Checkout stays responsive even when the payment provider is slow or retrying.
• Next.js web and React Native share one TypeScript contract. Booking models have a single source of truth, which kills a whole class of web/mobile integration bugs.
• RAG plus LangGraph agents (OpenAI and Claude, pgvector retrieval over routes and policies) handle multilingual support and itinerary generation without hand-writing copy for every market.
• Everything is code: Kubernetes on EKS, Terraform, and Ansible, wired through GitHub Actions for reproducible deploys across environments and markets.

• Multi-leg journey search and a seat-selection checkout with Adyen, live across Portugal, Spain and Italy.
• Full internationalization across 6 languages, both interface and generated content.
• Web and mobile from one codebase: a Next.js app and a React Native app sharing types and API contracts.
• An AI support and content layer: RAG pipelines and LangGraph agents for internal workflows, plus LLM itinerary generation and summarization.
• CI/CD and infrastructure-as-code: Terraform-provisioned EKS with GitHub Actions pipelines.

• Rebuilt the seat-selection checkout and lifted conversion by 15%.
• Scaled from the launch markets to 6 countries and 210 routes with no platform rewrite.
• Agent automation cut internal support handling time by 80%.
• Brought checkout p95 latency under 400 ms, roughly half its previous time.