I built an AI code reviewer that runs inside our VPN (Go and Bedrock) (alexandrecastro.tech)

A senior engineer's pull request (PR) nearly contained a critical bug, which was caught by Archbot, a custom AI code reviewer designed to operate within a VPN. Traditional AI code review tools were not an option due to privacy concerns, prompting the developer to create Archbot to analyze proprietary code securely. The initial implementation struggled with context limitations, leading to misidentified errors, but improvements followed, utilizing Repomix to gather necessary context efficiently and implementing a two-phase architecture that intelligently selected relevant files for AI review. Archbot's advancements included optimized query handling, structured output from AWS Bedrock Tools, and filtering of extraneous information to reduce token costs per review. By refining this process, reviews that previously cost approximately $0.20 each now average just $0.02, significantly lowering operational costs while improving review speed and accuracy. This case study highlights the potential of tailored AI solutions in enterprise environments, illustrating how effective tool integration and structured prompts can enhance code quality assurance processes.