Yeah, this gets messy fast once you start duplicating squads per language, you’re basically maintaining parallel systems. A cleaner approach is to keep a lightweight “router” layer upfront (language detection → route to a dedicated Vietnamese or English squad), instead of stuffing everything into one mega squad. That keeps logic isolated and easier to scale. Also, for Vietnamese, you might get more stable results using a dedicated STT model instead of multilingual fallback that inconsistency is a known pain point. I can help you set this up cleanly if you want , how are you currently handling language detection and routing today?