Property Finder reports 86 percent lower image AI costs after evaluation overhaul

Property Finder reports achieving higher image assessment accuracy at 86 percent lower cost after rebuilding its evaluation labels and selecting Llama 4 Maverick 17B Instruct on Amazon Bedrock, in a case study co-authored with AWS. The real estate platform processes tens of thousands of property images daily across the Middle East and North Africa. The team found that 96.8 percent of its evaluation labels came from its incumbent model's predictions. Fresh expert annotations enabled comparisons against human judgments. On 356 UAE images, the selected model scored 0.80 on Micro F1, a measure of label correctness, versus the incumbent's 0.56. However, the evaluation included 256 development images used for prompt tuning, making scores optimistic; the selected model also received more tuning effort.