Embed the world: Multimodal AI for searchable aerial imagery at scale

Short summary

AWS built a searchable aerial imagery system using Amazon Nova Multimodal Embeddings, demonstrating how multimodal AI enables semantic search at scale. The architecture evolved into Vexcel Intelligence and includes lessons on embedding models, fusion strategies, and evaluation methodology. Amazon Nova achieved highest F1 scores for geospatial semantic search.

• •Multimodal embeddings (Amazon Nova) outperformed other models for aerial image search
• •Architecture uses Amazon Bedrock and OpenSearch Serverless for scalable semantic search
• •Methodology and design patterns applicable to other geospatial and image-search products

Generated with AI, which can make mistakes.