Idea
An AI model aligning text and images of listings for improved search and recommendation in real estate and rental platforms.
Research Paper
Core Innovation
This paper introduces BiListing, which creates unified embeddings for both text and images of listings using large language and pretrained language-image models. Unlike prior approaches that treat modalities separately, BiListing enables cross-modal search and recommendation with a single embedding per listing. This approach improves zero-shot search capabilities and addresses cold start challenges effectively.
Market Size (TAM)
$20–50B TAM, $2–10B SAM; assumption: global online rental and real estate marketplaces expanding with AI-driven search improvements.
Potential Customers & Pain Points
- Online Rental Marketplaces Needing Better Search Relevance
- Real Estate Platforms Facing Cold Start Problems
- E-commerce Sites With Multimodal Listings
Business Model
Licensing the BiListing embedding API to online marketplaces and real estate platforms as a SaaS solution.
Competitive Landscape
- Zillow
- Redfin
- Airbnb
Implementation Challenges
- Integration with existing platform data pipelines
- Ensuring embedding quality across diverse listing types
- Scaling inference for large listing catalogs
Validation Strategy
- Pilot integration with a mid-size rental platform
- Measure search ranking improvements and user engagement
- Quantify incremental revenue impact post-deployment
Research Paper Overview
BiListing: Modality Alignment for Listings
Summary
BiListing aligns text and photos of listings by leveraging large-language models and pretrained language-image models to create a single embedding vector per listing and modality. This enables efficient zero-shot search, overcomes cold start problems, and supports listing-to-listing search across modalities. Deployed at Airbnb, it improved search ranking and generated significant incremental revenue.