VAR: Voyage Adaptive Reviews

VAR — Voyage Adaptive Reviews

A post-trip review system that reads what's already been said before deciding what to ask

expedia-var.vercel.app (generative aspect of questions won't work as hackathon openai api key has expired - placeholder questions have been used as a replacement for now)

Most review forms don't know what they don't know.

They ask the same questions to every traveler at every property. How was the staff? How was the room? It doesn't matter that staff has been mentioned in 1,100 of the last 6,000 reviews, or that wifi has come up exactly 36 times, or that nobody has commented on noise levels in over a year. The form asks anyway. The traveler sighs, clicks through, or doesn't bother.

VAR starts differently. Before generating a single question, it reads the existing review corpus for that property and computes which topics are over-covered, which are absent, and which were relevant a year ago but haven't surfaced since. Low submission rates aren't a marketing problem. They're a design problem. If a review feels like homework tacked onto the end of a trip, people skip it. So VAR embeds review inside something the traveler already wants: a replay of where they just went.

The result is two questions. One targets the highest-priority blind spot in that property's review history. One verifies a specific structured claim. From those two answers, a three-stage AI pipeline synthesizes a complete review with a title, body, and per-category ratings, written back into the database and immediately usable for discovery and ranking.

Ask what matters. Ask it once. Make it feel like the end of the trip.

Built in 72 hours for the 2026 Wharton AI & Analytics Hack-AI-Thon, presented by Expedia.

How It Works

1. The system ingests existing reviews for a property and computes a coverage vector of which topics are missing and which are stale, weighted by a 1-year half-life decay so recent reviews count more than old ones.
2. Stage 2 generates one gap question targeting the blind spot and one verification question, with an explicit reasoning field explaining why this question, why now, for this property.
3. The user answers via voice or text inside a cinematic JFK to destination property experience, so the review never feels like a form.
4. Stage 3 synthesizes both answers into a full Expedia-style review with inferred per-category ratings, persisted to Supabase and surfaced back to the user as the trip's closing beat.

No static prompts. No redundant questions. Just signal that the property is actually missing.

Technical Innovation:

The system demonstrates advanced natural language processing capabilities including:

• Review Corpus Analysis - Semantic analysis of existing reviews to identify coverage gaps
• Topic Modeling - Computing topic relevance and decay over time
• Question Generation - AI-powered synthesis of targeted, context-aware questions
• Multi-modal Input - Voice and text input capabilities
• Real-time Synthesis - Three-stage pipeline for immediate review generation

Impact:

VAR transforms the traditional review process from a repetitive form-filling exercise into a meaningful conversation. By understanding what's already known, the system asks only what matters, making the review process more efficient and the resulting content more valuable for future travelers.

This project represents a breakthrough in user experience design, showing how AI can eliminate redundant interactions and focus on genuine information gaps.