Halodoc built an AI-powered code review system embedded in their Jenkins Global Library CI/CD pipeline, reducing Merge Request feedback time from 30–45 minutes to 1–2 minutes. The system uses two LLMs — Gemini CLI for fast diff-based reviews and Claude Sonnet via AWS Bedrock for deeper contextual analysis — with stack-specific prompts covering Java, Angular, Swift, Kotlin, SRE, and Data Engineering repositories. A Comment Lifecycle Manager prevents duplicate discussions and auto-resolves fixed findings. Over six months, the platform processed 58,811 review invocations (~9,800/month) at roughly $1,600–$1,700/month in inference costs, with 85% of findings accepted by developers. Key lessons: prompt engineering and guardrails matter more than model selection, and lifecycle management is essential for developer trust.
