Andy Wong has worked on Rovo Dev Code Reviewer at Atlassian, helping develop systems that select which AI-generated pull-request comments reach developers. His collaborative engineering and research connects a specific design decision—separating feedback generation from feedback selection—with deployment evidence about which comments lead developers to change code.
Andy’s background includes AI development and integration and large data platforms. He studied at Western University from 2014 to 2018. Atlassian’s April 2026 engineering account, which he co-authored, identifies him as a software engineer working on its automated reviewer.
Choosing which comments to show
Andy co-authored Atlassian’s August 2025 account of the comment ranker, developed with colleagues in the DevAI organization. Early internal use exposed comments that were noisy, overly particular, or wrong. The team initially used heuristic filters, including selecting certain categories of comments, before developing a learned model trained on developers’ responses.
The ranker separates comment generation and selection. A language model proposes feedback using the code change and associated work descriptions; a classifier then predicts which suggestions are likely to prompt a code change. Its training signal is code resolution: whether a developer subsequently changes the line associated with a comment. This provides feedback from actual use, with a limitation: a changed line is a proxy for usefulness, rather than proof that the suggestion was correct.
The team reported that the ranker continued to work as the underlying generator changed from GPT-4o to Claude Sonnet models. Separating selection from generation allowed the team to change the model producing comments while retaining a filter informed by developer behavior.
Evaluating review in everyday development
Andy also co-authored the large-scale evaluation of RovoDev Code Reviewer, a collaboration involving Atlassian engineers and Monash University researcher Kla Tantithamthavorn. Across a year of deployment in more than 1,900 internal repositories, the study examined over 54,000 generated comments. It reported that 38.7% led to subsequent code changes and observed a 30.8% reduction in median pull-request cycle time. Those findings describe the team’s system in Atlassian’s development environment; they do not establish that other teams will see the same improvements.
In the April 2026 engineering account, the agent supplies an initial review pass while engineers decide which suggestions to accept. The ranker addresses one part of that arrangement: reducing the comments developers must assess while preserving feedback they may act on.
Turning team standards into review rules
In The SDLC You Build Yourself, presented with Andrei Bocan and Shane Wolf under an Atlassian affiliation, Andy demonstrated how a design token standard becomes a reusable code-review rule. The workflow lets a team select the repositories where the standard applies, rather than enforce it by hand on every pull request. This adds an explicit team requirement to the review process alongside the learned selection of useful comments.
Andy Wong, Andrei Bocan and Shane Wolf explain Atlassian’s approach to connecting Jira work, organizational context, repository standards and autonomous development—with human judgment at the points where requirements, evidence and review capacity matter.
The context engine combines connected organizational information with code search, so agents can retrieve relevant implementation and decisions across separate repositories.
Fresh retrieval, human verification and historical reasoning serve different purposes: staying current, choosing trusted sources and understanding why the code took its present shape.
Autonomous development needs separate controls for spending and human review capacity, plus explicit pauses for missing requirements and completed changes awaiting judgment.
Acceptance criteria become actionable when the agent also has a runnable environment and repository instructions that explain how to verify the result.
▶ Watch ↗