Sentry has implemented an AI-powered solution to improve its issue grouping algorithm, reducing new issues created by 40% while maintaining sub-100ms end-to-end processing latency. The new approach uses a transformer-based text embeddings model to capture the semantic essence of each error, outperforming simplistic edit-distance measures in recognizing similar errors. The system was tested extensively with thousands of example issues and showed virtually zero false positive issue grouping rate, while providing significant reduction in incorrect new issue creation. To optimize performance, Sentry used techniques such as memory right-sizing, hash partitioning, and efficient query implementations, ultimately enabling the AI-powered solution to be deployed at scale without sacrificing performance. The change was rolled out slowly over several months, with a 40% drop in new issues created observed overall, making it available to all Sentry users at no additional cost.