Gazelle: Smart validation and 99.9%-accurate property descriptions with Gemini

Overview

Gazelle, a Swedish startup automating property documentation for real estate agents in Sweden and Norway, uses Google's Gemini 1.5 Pro to extract key information from lengthy property documents like building-inspection and electricity reports and generate sales content. Gemini increased output accuracy from less than 95% to 99.9%, reduced content generation time from four hours to 10 seconds, and helped Gazelle launch four new products in under a year, saving brokers 4 to 6 hours of manual work per property.

The challenge

Extracting key information from lengthy property documents was a manual, time-consuming task for real estate brokers; Gazelle found certain documents too complex for traditional machine-learning methods, and its third-party generative AI tool would hallucinate answers, which was unacceptable since customers could get sued for presenting incorrect information.

The solution

Gazelle adopted Gemini 1.5 Pro from Google Cloud to extract and summarize key information from lengthy documents like building-inspection and electricity reports, using Gemini's multimodal capabilities to recognize information across document types and images, integrating via APIs directly into real estate agents' software, and using BigQuery to analyze tool performance; it is also developing a Google Maps Platform and Google Search-grounded feature to generate area descriptions and image captions.

Reported business value

Gazelle increased the accuracy of its output from less than 95% to 99.9% with Gemini, cut content generation from four hours of manual work to 10 seconds, and saves brokers 4 to 6 hours of manual work per object; faster quality control also let Gazelle launch four new products in under a year, and Gemini costs around a fifth of the price of alternatives.