Problem & Challenge
Traditional travel agencies often operate through fragmented communication channels: customer discovery occurs through manual chats, tour pricing relies on unstructured spreadsheets, and package curation requires days of back-and-forth consultation with airline and hotel supplier portals.
FlightOne was conceptualized as an integrated AI travel platform designed to automate this lifecycle. The platform needed to fulfill two critical mandates simultaneously: deliver a bespoke web portal that showcases international tours (Maldives, Dubai, Turkey, Thailand), while providing a dedicated conversational AI consultant (Ava) capable of flight search, hotel tier pricing (3★, 4★, 5★), and booking orchestration.
Technical Approach & Architecture
FlightOne is engineered as three tightly decoupled yet unified systems:
- Agency Web Portal: A high-performance Next.js web application built with responsive SCSS modules, Lenis smooth scrolling, and dynamic destination visualizers.
- Conversational Assistant Client: A lightweight React application where travelers converse with Ava, explore structured recommendations, refine dates, and review flight fares.
- Modular API Server: An Express.js backend using Prisma ORM on PostgreSQL that enforces deterministic pricing rules, manages itinerary drafts and handles token authentication.
Engineering Challenges & Solutions
1. Deterministic Pricing vs. Generative LLM Variance
Language models are probabilistic and can produce wrong numbers when calculating flight fares, currency conversions, and seasonal surcharges. A price invented by the model would be a real financial error.
Solution: We separated semantic understanding from business calculations. Ava handles natural language intent extraction and structures parameters into typed JSON payloads. The Express/Prisma server applies the pricing rules and tiered markups and returns every fare to the client, so the model never calculates or invents a price.
2. Cross-Device Consultation Continuity
Travelers frequently initiate trip consultations on mobile devices but complete itinerary reviews on desktop computers. Maintaining active session state without requiring mandatory upfront registration was essential for user retention.
Solution: Implemented guest UUID tokens stored in encrypted client cookies that bind conversation threads to ephemeral PostgreSQL session records. When travelers choose to register or authenticate via JWT, the server merges the guest consultation history into their permanent profile seamlessly.
Key Features & Capabilities
- Bespoke 24-Hour Itinerary Pipeline: Automated day-by-day travel plan generation tailored to honeymooners, families, and corporate groups.
- 3-Tier Accommodation Benchmarks: Instant comparison matrices between 3-star budget, 4-star boutique, and 5-star luxury options.
- Multi-Repo Architecture: Clean domain isolation across agency storefront, client chat UI, and server backend.
- Human Escalation Fail-Safe: Seamless handoff from AI consultant to human travel agents for visa-sensitive approvals.