
THE PROBLEM A salon owner in Pune misses 30% of her calls while she's mid-haircut. Every missed call is a lost booking - and in India, customers don't leave voicemail, they call the next salon. Callers don't speak in neat English sentences either. They speak Hinglish: "kal shaam ko slot milega kya?" WHAT IRA DOES Ira picks up on the first ring and talks the way the caller talks - Hindi, English, or the code-mixed Hinglish 600+ million Indians actually use. She checks a live, duration-aware slot book (a 60-minute facial blocks both half-hour cells), offers the nearest alternative when a slot is taken, reads the booking back before writing anything, and sends a Hinglish WhatsApp confirmation. Call again from the same number and she recognises you - reschedule or cancel in one call. She is honest by construction. Ask for a hair spa when the catalog has haircut and facial, and she says exactly that instead of inventing a service. She only books slots the engine returned and only quotes catalog prices. TECH - AssemblyAI realtime streaming STT (universal-3.5-pro), turn detection tuned for phone cadence, keyterm hints for salon vocabulary, automatic language tracking across Hindi/English code-switches. Mic streams PCM 16kHz over secure WebSocket; barge-in supported. - Tool-calling LLM brain (OpenAI-compatible, swappable): the model never touches the booking store directly; it calls engine functions and speaks the results. - Duration-aware booking engine with JSON persistence: one active booking per phone number. - WhatsApp Cloud API confirmations (public demo runs in demo mode). - Node 20 / Deno, Docker, MIT licensed. WHY IT WINS India's real phone traffic is code-mixed - Ira is built for the calls that actually happen. A missed call is measurable lost revenue for 60M+ small service businesses; Ira converts it into a confirmed booking. Realtime code-switch STT, barge-in turn-taking, grounded tool calls - not a prompt wrapper.
30 Sep 2026