Unpopular opinion: when building a AI SaaS, in-house AI based implementation is much better compared to using external libraries like llama index. For smaller use cases it works well, but for complex workflows it often fails, mostly because AI development is really fast, interconnected dependencies cause problem in production that will slow down engineering and hinder user experience.