Amazon’s catalog is huge, messy, and constantly changing—and that is exactly why the team moved to a self-learning setup built to digest millions of submissions every day. Instead of treating model disagreements as something to patch over, they treat them as moments worth paying attention to: a signal that the system has found an edge case, a messy description, or a term it does not truly understand yet. Amazon Bedrock sits in the middle, coordinating a layered “workforce” of Foundation Model instances. Most of the time, the work is handled by several smaller “worker” models—cost-efficient options such as Amazon Nova Lite—running in parallel to extract product attributes. If the workers line up on the same answer, the system moves on. When they do not, or when the input is simply ambiguous, a stronger supervisor AI Agent—powered by Anthropic Claude Sonnet—steps in to take a closer look. Crucially, the supervisor is not there just to provide a one-off fix. It tries to understand why the workers split, and it writes down the lesson in a form the system can reuse. Those lessons are organized in a hierarchical knowledge base and fed back into the worker layer through Prompt Engineering. Over time, this feedback loop helps the LLM get better at the thorny, domain-specific language that shows up in real product data. The practical payoff is straightforward: spend expensive compute only where it matters most, and let everyday throughput stay fast and cheap—while accuracy gradually improves.

Amazon’s catalog is huge, messy, and constantly changing—and that is exactly why the team moved to a self-learning setup built to digest millions of submissions every day. Instead of treating model disagreements as something to patch over, they treat them as moments worth paying attention to: a signal that the system has found an edge case, a messy description, or a term it does not truly understand yet. Amazon Bedrock sits in the middle, coordinating a layered “workforce” of Foundation Model instances. Most of the time, the work is handled by several smaller “worker” models—cost-efficient options such as Amazon Nova Lite—running in parallel to extract product attributes. If the workers line up on the same answer, the system moves on. When they do not, or when the input is simply ambiguous, a stronger supervisor AI Agent—powered by Anthropic Claude Sonnet—steps in to take a closer look. Crucially, the supervisor is not there just to provide a one-off fix. It tries to understand why the workers split, and it writes down the lesson in a form the system can reuse. Those lessons are organized in a hierarchical knowledge base and fed back into the worker layer through Prompt Engineering. Over time, this feedback loop helps the LLM get better at the thorny, domain-specific language that shows up in real product data. The practical payoff is straightforward: spend expensive compute only where it matters most, and let everyday throughput stay fast and cheap—while accuracy gradually improves.