Property Finder replaced model-derived labels with expert annotations and selected Llama 4 Maverick 17B Instruct for image checks on Amazon Bedrock.
Property Finder says it chose Meta’s Llama 4 Maverick 17B Instruct for its property-image quality pipeline after finding that its existing evaluation set largely reflected the model it was meant to assess. In an AWS post published October 27, 2026, the teams report that 96.8% of the labels had come directly from predictions by the incumbent vision-language model rather than independent expert review.
The real-estate platform processes tens of thousands of images each day across the Middle East and North Africa. Its checks cover 11 attributes, including room type, furnishing, watermarks, image quality and construction indicators. To create a more independent comparison, the team had experts annotate 356 UAE images without seeing earlier model outputs; 256 of those images formed the development set. A separate Egypt sample was used for a cross-regional check, but the post does not report its results.
The AWS Generative AI Innovation Center and Property Finder iterated on prompts, addressing issues such as rooms visible through doorways, Arabic decorative text that could be mistaken for watermarks, and output formatting. The teams used evolutionary prompt optimization, and the final prompt included additional Property Finder rules for renders, image borders and corridors.
On the 356-image evaluation set, Llama 4 Maverick 17B Instruct with the final prompt reached a Micro F1 score of 0.80, compared with 0.56 for the incumbent model using its prompt. The teams caution that the evaluation set includes development images used for tuning, so those results are optimistic. Property Finder reported that the selected model matched or exceeded the incumbent on all 11 attributes; the two tied on multi-label photo classification.
AWS estimates the new pipeline at $2.35 per 1,000 images, an 86% reduction against the prior setup’s cost estimate. Those cost figures are based on December 2025 prices in the US East (Ohio) Region and specified token assumptions. Property Finder also observed 40–60% lower inference latency in its own production monitoring. The company runs the pipeline on Amazon Bedrock and selected the model based on accuracy across its attributes and cost efficiency.