In this post, we dive deep into the architecture and techniques we used to improve Miro’s bug routing, achieving six times fewer team reassignments and five times shorter time-to-resolution powered by Amazon Bedrock.

Automating bug triaging in modern software environments is complex. Bug reports are often messy, lack context, and contain diverse data including text, stack traces, screenshots, and even videos. The complexity multiplies in software-focused companies with many teams, creating a multi-class classification problem with numerous possible labels. Miro’s engineering organization consists of nearly 100 teams, each responsible for specific product aspects.High-accuracy bug classification requires augmenting reports with relevant product information from various sources, including GitHub pull requests, Confluence documentation, README files, and previously resolved tickets. Additionally, organizational structures are dynamic—teams merge, new teams form, and products evolve, continuously changing team responsibilities. The same holds true for software features that are added, updated, or deprecated.Traditional natural language processing (NLP) based text classifiers, such as fine-tuned BERT models or fine-tuned large language model (LLM) classifiers, face severe limitations in these dynamic environments. They require retraining when organizational changes occur and depend on labeled data that might not exist for new structures. Miro experienced quickly degrading performance with an existing solution based on a fine-tuned GPT model.Recognizing these challenges, Miro opted for an LLM-powered approach that combines optimized prompts for team classification with Retrieval Augmented Generation (RAG) for context retrieval, creating a more adaptable, zero-training, and higher-accuracy bug triaging solution: BugManager.

BugManager adopts an LLM-powered approach to bug classification. When a new bug report is received, the BugManager classification system takes action. BugManager first parses non-text data such as screenshots or screen recordings using the multimodal image and video understanding capabilities of Amazon Nova Pro. The system then enriches the parsed report with important context from several knowledge bases (using Amazon Bedrock Knowledge Bases) with RAG. These knowledge bases contain, for example, previously resolved Jira issues, GitHub pull requests, Confluence documentation, and GitHub READMEs. Using Anthropic’s Claude Sonnet 4 on Amazon Bedrock, the system combines the enriched bug descriptions along with detailed textual information on each team and their responsibilities into a single, optimized classification prompt that performs the routing to the correct team. As an optional feature, the system can also generate a detailed root cause analysis, using the collected information in combination with retrieval of relevant source code repositories to provide deeper insights into the issue and offer hypotheses on how to resolve them.

To enable frictionless adoption and use, BugManager is exposed to the Miro Engineering community through a simple Slack-based workflow. The following screenshot shows an example of a user interaction with BugManager. Based on the initial description of the bug, BugManager proposes up to five suitable teams (in order of priority) along with a rationale. By default, the bug is routed to the most likely team, but users can manually overwrite this selection. Optionally, BugManager also provides engineers with a root cause analysis of the reported bug, drawing on information retrieved from the full Miro code base.