Garage Eight is integrating an AI-based code reviewer to automate the merge request (MR) approval process. In October, iOS teams removed the third mandatory approval for MRs, limiting the requirement to a maximum of two. In the internal analytics service, 69 MRs were merged without any human approval.
The company classifies services into four tiers based on the potential impact of errors. Tier 3 and 4 services, where errors do not affect users, are the primary focus for removing human reviewers. As of mid-October, 41% of microservices have been assigned a tier by architects.
The AI reviewer assesses the complexity of each MR, labeling them as simple, normal, or complex. For Tier 1 and 2, which involve user-facing functions or financial operations, human oversight remains mandatory. The company uses different models depending on the tier, prioritizing cost-efficiency for lower-risk code.
Effectiveness is measured by code changes following AI feedback rather than just resolution rates. In the third quarter, 37% of major and critical findings led to code modifications. The company aims to maintain a 'not useful' feedback rate below 10% for AI suggestions.
Future plans include implementing an automated 'fuse' that restores mandatory human approvals if rollback rates increase. The team continues to refine the system by checking against a checklist of seven criteria, including tier classification, evaluation sets, and alert systems for API usage.
