Distributed backend

The app needed ordinary operations (users, content, tickets) and AI work (itineraries, review summaries). The AI work is slower, changes often and uses different tools, and it shouldn't interfere with the core operations.

I split the backend into two services for isolation. The core API is written in Node.js and Express. The AI service is in FastAPI and Python, and more machine-learning work can be added there later. The core service is the main component: the mobile app talks only to it, and it manages the AI service. A request that needs AI goes mobile app, core service, AI service. The two services talk with authenticated machine-to-machine calls, and the API keys rotate automatically.

Two practical reasons backed the split:

• Tooling. At the time, LangChain and LangGraph were not fully supported in JavaScript.
• Familiarity. Fewer people on the team were comfortable with Python. Separate services let the people who know Python work on the AI piece, and the JavaScript people keep working on the backend without any problem.

A single service was the first thought. I insisted on separating them because of the isolation and the tooling.

Architected a distributed backend consisting of a core API service and a dedicated AI service, separating standard application functionality from AI-specific workloads.

The core service acts as the primary backend interface for the mobile application. Requests requiring AI functionality are routed through the core service to the dedicated AI service:

Mobile Application → Core Service → AI Service

Implemented secure machine-to-machine communication between the services, including authentication and automated API-key rotation.

This architecture provided a clear separation between core application responsibilities and AI workloads while maintaining controlled communication between the services.

• A network hop. Every AI request goes from the core layer to the AI layer, which adds a hop.
• Key rotation. The keys between the services rotate automatically every six months, which is one more process to keep running.

The application and AI work can be developed and operated independently, and each team works in the language it knows.