The case study
you can
interrupt.

What breaks in conventional systems

• Language is chosen once, up front, and enforced for the whole call.
• Code-switching mid-sentence confuses single-language ASR into garbage transcripts.
• Latency over ~1.5 seconds reads as "it hung up on me" and callers start talking over the system.
• No barge-in: the system finishes its paragraph while the caller repeats "agent, agent, agent".
• Silence during connection and transfer reads as a dropped call.

What the reference build does instead

• Detects language per utterance, not per call — the agent follows the caller.
• Streams ASR → LLM → TTS so first audio lands in under a second on our bench.
• Stops speaking the instant the caller starts (barge-in), then recovers the turn cleanly.
• Plays deliberate sound cues — ringback, connect, hangup — so the line never feels dead.
• Escalates to a human with full transcript context when confidence drops.

The check that runs on every message

// Script detection — from the shipped demo
const DEVANAGARI = /[\u0900-\u097F]/; // हिन्दी
const TAMIL      = /[\u0B80-\u0BFF]/; // தமிழ்

function detectLang(text: string) {
  if (DEVANAGARI.test(text)) return 'hi';
  if (TAMIL.test(text))      return 'ta';
  return 'en';
}

Why it matters commercially

• Callers stay in their strongest language, so resolution doesn't depend on their English.
• Romanized Hindi ("mera card block ho gaya") is handled by the agent's multilingual model, not the script check — the two layers cover each other.
• The transcript stores language per turn, which makes QA and analytics honest about what the caller actually said.
• Adding a language is configuration plus evals, not a rebuild.