Miro's BugManager on Amazon Bedrock cuts bug routing errors by 6x, resolution time by 5x
Miro's AI bug triager recovers 42 years of lost developer productivity annually.
Miro, an AI-powered innovation workspace with over 95 million users, faced a chronic bug triaging problem: misrouted tickets across nearly 100 engineering teams caused an estimated 42 years of cumulative lost productivity annually due to reassignments and redundant investigations. To solve this, Miro partnered with AWS's PACE team to build BugManager, an LLM-powered solution on Amazon Bedrock.
BugManager uses a RAG (Retrieval Augmented Generation) architecture. First, it parses multimodal inputs like screenshots and screen recordings using Amazon Nova Pro. Then it enriches bug reports with context from Amazon Bedrock Knowledge Bases containing past Jira issues, GitHub PRs, Confluence docs, and README files. Finally, Anthropic's Claude Sonnet 4 on Amazon Bedrock classifies the bug to the correct team using optimized prompts and detailed team responsibility descriptions. The system also optionally generates root cause analysis by retrieving relevant source code. This zero-training approach eliminates the need for retraining when teams or features change, achieving 6 times fewer reassignments and 5 times shorter time-to-resolution compared to Miro's previous fine-tuned GPT model.
- BugManager combines Amazon Nova Pro for multimodal parsing, RAG from Bedrock Knowledge Bases, and Claude Sonnet 4 for classification.
- Miro reduced bug misrouting by 6x and cut resolution time by 5x, recovering 42 years of annual lost productivity.
- The zero-training RAG approach adapts to dynamic team changes without retraining, unlike fine-tuned NLP classifiers.
Why It Matters
AI-powered bug triaging can dramatically cut developer toil and speed up product fixes at scale.