Author
Rohan Mehta
Rohan owns AIVA's voice pipeline — the system that takes a phone call from ring to reply in under 200ms. Previously built real-time systems at Razorpay.
Posts by Rohan Mehta
Voice AI response latency: why we target 300ms
We didn't pick 200ms because it sounded fast. We picked it because it's under the gap two humans leave each other in conversation — and callers can tell.
AI agent for accountants: absorb the tax season spike
For six weeks a year your phone doesn't stop, and the other 46 you don't need the headcount. An AI agent absorbs the spike without a seasonal hire.
AI agent for courier and logistics pickup requests
Order volume jumps hard during a festival sale, and so does the phone. An AI agent takes the pickup request the first time, no call center hire.
AI agent for real estate: never miss a buyer's call again
A buyer who hits voicemail on a Sunday won't call back — they'll call the next number on the portal. AIVA picks up in under 200ms, every time.
AI phone agent for restaurants: the dinner-rush call
The phone rings loudest at exactly the moment nobody has a free hand — 7:45pm, kitchen slammed, host seating a table. AIVA was built for that minute.
AI receptionist for auto repair shops: booking bay slots
A customer describing a grinding noise mid-drive is a bad candidate for hold music. An AI receptionist for auto repair shops picks up and books the bay slot.
AIVA + Twilio: what's actually happening under the hood
Our integrations page reduces Twilio to one line: powers voice and SMS, with failover. Here's the actual division of labor between Twilio and AIVA's own stack.
Frankfurt and Virginia: why we run inference outside India.
Voice under 200ms becomes a physics problem once customers aren't in India. Why we stood up inference in Frankfurt and Virginia — and what it costs.
Interruptions and pauses in natural voice AI conversations
Real people interrupt, backtrack, and correct themselves mid-sentence. A voice agent that can't keep up isn't slow — it's having a different conversation.
Routing calls to regional inference: Mumbai to Virginia.
You can optimize a model all you want. You can't optimize away the distance between Mumbai and Virginia. So we put the inference where the call actually is.
How we cut AI voice latency from 240ms to 198ms (2026)
Inference pipeline rewrite, regional caching, and a Twilio WebRTC encoder setting nobody documents — how we took AI voice latency from 240ms down to 198ms.
Rewriting the voice pipeline (and why we'd do it again).
Five vendor hops, a 380ms p95 we couldn't move, and a six-week parallel run. What we learned rebuilding the AIVA voice stack from scratch — and what it cost.