We are using judges with LLMs and web grounding plus manual grading. We recently did a benchmark on the LLM quality across major AI providers - we plan to open source it soon and will probably open source our API quality check benchmark too https://news.ycombinator.com/item?id=47366423