MCP Server Integration & Voice Assistant MVP Development (1-2 Week Sprint)

Senior AI & Backend Engineer – Airline MCP Server Integration & Voice Assistant MVP Development (2-Week Sprint) Project Overview: As part of an airline's innovation program, we are building a voice passenger assistant prototype (MVP) to be deployed at the airport. We have a dynamic 2-week development sprint ahead of us to reach the final submission deadline. Critical Structural Note: You will NOT build backend logic or a database from scratch for this project. Your core responsibility is to build an MCP Client (Model Context Protocol) that connects to the airline's already active remote MCP Server and triggers its existing tools/functions via voice commands. Technical Stack & Architecture (MVP Phase): • Backend: Node.js or Python (Fully containerized and configured to spin up with a single docker-compose up command). • AI & Cloud API Layer: o STT (Speech-to-Text): Cloud-based Whisper API integration combined with browser-based audio capture (Web Audio API) for low-latency transcription. o LLM Orchestration: Anthropic Claude 3.5 Sonnet API. Since the airline's remote MCP server is natively optimized for Anthropic's tool-calling architecture, Claude 3.5 will be used for rapid integration during this MVP phase. • Frontend: HTML5, CSS3, and Modern JS (React, Vue, or Vanilla JS). The application must be fully responsive, working as a full-screen app on mobile devices (PWA-ready) and rendering inside an elegant smartphone mock frame on desktop browsers. MVP Core Engineering Features to Implement: 1. Acoustic & Cognitive Filter Layer (Intent Analysis): Airport environments are highly noisy, and passengers often speak in broken, panicked, or fragmented sentences. • The raw text coming from the Whisper API must NEVER be fed directly into the core system. • Claude 3.5 Sonnet must intercept the raw text, filter out semantic noise, and output a clean "Context Summary" card on the UI (e.g., "What I understood: You want to claim a meal voucher for your delayed flight to Paris."). • The model must strictly enforce tool calling / structured outputs (Strict JSON Mode) to generate the precise JSON schemas required by the airline's MCP server. 2. Multi-Tool Chaining: The developed client must orchestrate the following core passenger workflows based on a single voice prompt by interacting with the airline's remote MCP tools: • Voice Check-In: Dynamically load the seat map and trigger an Apple/Google Wallet push notification upon confirmation. • Gate Info & Navigation: Fetch real-time flight status and visualize terminal walking routes. • Lost Baggage Reporting: Parse the passenger's verbal description of missing luggage and auto-fill the official PIR form. • Loyalty Program Actions: Retrieve the passenger's mileage balance and present a one-click confirmation button to upgrade to Business Class. • Disruption/Delay Management: Detect delayed flights and dynamically display a QR-coded digital dining voucher on the passenger's screen. 3. Graceful Out-of-Scope Routing: If the user asks questions completely unrelated to airline operations (e.g., "Write me a poem"), the orchestrator must prevent hallucinations or system failure. It must gracefully output a polite corporate message routing the user back to terminal features or opening the static city-guide page. UI/UX Design Note: Frontend development, including encoding screen animations (voice waves, flip-board loading effects) and user experience, is INCLUDED in this job scope. However, to avoid technical clutter, an extremely detailed UI/UX Design Guideline—containing corporate color codes, logo placements, screen layouts, and user flowcharts—will be shared separately during the onboarding phase. Future / Production Vision Requirement: All code must be completely Model-Agnostic and OpenAI-compatible. Although the cloud-based Claude API is used for speed during the MVP phase, the next phase will involve running the system completely offline on local servers (On-Premises) using open-source models like Qwen 2.5 32B. Therefore, the backend architecture must be highly modular so that the endpoint (Base URL) can later be swapped out for a local inference engine (vLLM) without breaking any core logic. Deliverables: 1. A clean Git repository containing frontend and backend code that spins up locally with a single docker-compose up command. 2. A comprehensive README.md file detailing the MCP connection parameters and setup instructions. 3. A 15-minute technical walkthrough (Loom) video explaining the code architecture and multi-tool chaining logic.