Hello @Rohan ! I saw your post about optimizing the VAPI interview assistant. I'm a senior engineer with extensive hands-on experience building and tuning real-time voice agents (STT->LLM->TTS pipelines) in production. The cut-off vs slow response issue is typically a turn-detection and streaming orchestration problem - not just a provider issue. In past systems, I've reduced perceived latency below 2s by: - Using fully streaming STT + LLM - Triggering generation only after stabilized partial transcripts - Tuning silence thresholds (~800–1000ms depending on cadence) - Streaming TTS on early token batches - Separating barge-in logic from endpointing logic I'd be very interested in contributing long-term and helping you make the assistant feel truly conversational.