使用 LLM 作为评估者评估 LLM 代码生成

This tutorial demonstrates how to score every code generation response against custom criteria, enabling teams to select the AI model that best meets their security requirements, API schemas, and blind spots. By setting up custom judges that check for vulnerabilities of actual concern, validate against real API conventions, and flag common scope creep patterns, teams can gather evidence over a few weeks to determine which model is suitable for which tasks. The tutorial involves building a proxy server that routes Claude Code requests through LaunchDarkly, allowing requests to be forwarded to any model, including Anthropic, OpenAI, Mistral, or local Ollama instances.