Orca Security cuts cloud query costs by 90% with Amazon Bedrock agent

The system uses a nine-stage pipeline with specialized models—including Anthropic Claude Haiku 4.5 for reasoning and Amazon Titan for embeddings—to parse user requests, filter irrelevant models, and validate outputs. A human-in-the-loop labeling system continuously improves accuracy by letting domain experts correct misclassified queries. Orca’s Sonar platform supports 900+ cloud models, but the agent dynamically narrows context to only relevant schemas, avoiding the inefficiency of loading all definitions at once.