PSTN calls over Twilio and Exotel with Vapi orchestrating STT → LLM → TTS, call events over webhooks, voicemail detection so the agent never interviews an answering machine.

Callers can cut in, but a cough or an “hmm” doesn't kill the sentence: a false interruption resumes the reply after 1 s. Fixed cut-off speech on Hindi calls.

Keyterm prompting: product names, people and tech words are pushed into Deepgram nova-3 keyterms and Sarvam's prompt, so “InCruiter” stops coming back as something else.

STT auto-detects the language; the voice is picked from the reply itself (Hindi and Marathi share Devanagari, so marker words decide), and each language has its own TTS chain.

Script-aware, feature-flagged matching after an English-only guard scored a correct Devanagari answer 0.39 against a 0.5 threshold. Diagnosed Marathi STT hallucinating in 10 of 15 turns.

Every stage has a fallback chain: three STT engines, three LLMs, a TTS chain per language. A rate limit mid-call moves to the next model and the call carries on.

Per-stage latency budgets, streaming everywhere, preemptive generation that starts the reply before the turn is final, one 24 kHz format across voices.

Voice-specific prompting: hard word cap with a server-side trim, markdown and emoji stripped before TTS, scope lock, prompt-injection resistance, no invented facts, AI disclosure.

Krisp noise cancellation on the caller's audio in the agent, echo cancellation and auto gain in the browser.

One-minute mic-only tokens bound to this site's origin, per-IP rate limit, 5-minute cap, hang-up after 45 s of silence, no call recording.