Back to Projects
A
AI/ML

ArSphere

Full-Stack & AI Engineer

React NativeNestJSFastAPILangChainOpenAIAnthropicGroq

About the Project

ArSphère is a luxury concierge platform where members interact with an AI concierge that arranges premium experiences in Paris — restaurants, shopping, hotels, chauffeurs, clefs d'or, and art and culture. It runs as three services across three runtimes: a React Native mobile client, a NestJS API, and a Python FastAPI service that owns the agent, speech and emotion layers. Membership is tiered and gated, and the design brief was specific: the concierge is not a search box. She must grasp a request about style, proportion and occasion, ask at most one clarifying question, and offer at most two options.

What It's Made Of

Mobile client

React Native · Expo 55 · Zustand

The only consumer surface — there is no web app. File-based routing with screens re-exported from feature modules, three Zustand stores for auth, chat and orders, and an Axios client that silently refreshes on a 401. Styling is deliberately StyleSheet-based rather than utility classes, because the className approach silently fails on gradients, image sizing and several third-party components. Voice input, the animated concierge avatar and the live chat all live here.

API

NestJS 11 · PostgreSQL · TypeORM · Redis

Owns identity, membership tiers, the offers catalogue, orders and bookings. A global JWT guard covers every route with explicit opt-out, and a membership guard sits in front of anything that would commit a real booking. Access tokens are short-lived with refresh in an httpOnly cookie; unknown DTO fields are stripped globally rather than validated per handler. It forwards every AI turn to the Python service and relays the result over Socket.io.

AI service — the concierge

FastAPI · LangChain · OpenAI · Anthropic · Groq

The agent itself, built on LangChain's chat-model abstractions so OpenAI, Anthropic and Groq sit behind one interface chosen by a factory — the model can be swapped, or failed over, without touching agent logic. It also runs the speech layer, an emotion parser so replies carry affect, and a viseme service that converts synthesized speech into the mouth shapes driving avatar lip-sync.

Key Highlights

  • Architected a three-service, three-runtime system: React Native mobile client, NestJS API, and a Python FastAPI AI service, sharing PostgreSQL and coordinating over HTTP and Socket.io
  • Built the concierge as a tool-calling agent on LangChain's chat-model abstractions — langchain-openai, langchain-anthropic and langchain-groq behind one interface with a factory, so the model can be swapped without touching agent logic
  • Implemented an emotion parser so responses carry affect, and a viseme service that drives avatar lip-sync from synthesized speech
  • Added full voice interaction with speech-to-text and synthesis, alongside real-time Socket.io messaging
  • Designed tier gating where a pending member sees the full value of a request but is elegantly blocked before any real booking — no 'confirmed' language is ever shown to an unverified account

Technical Challenges

The constraint that shaped the product was conversational restraint. A luxury concierge that interrogates you feels like a form, so the agent is capped at one clarifying question and two options — which means the retrieval and tool selection have to be right the first time, with far less room to recover through follow-up than a normal assistant has. Coordinating three runtimes around shared auth and real-time state also required strict API contracts, and the provider factory existed so an outage at one LLM vendor degrades the concierge rather than taking it offline.