Real-Time Voice AI Platform
Voice agents that call back customers who abandoned a booking, talk to them naturally in their language, and help them finish the purchase. They run over phone calls and in the browser.
Built at EaseMyTrip. The interface shown is an illustrative mock-up; internal details are generalised.
The problem
Many travellers get as far as the payment page and then leave. Calling every one of them back by hand is slow and expensive, and the callback often comes after the customer has lost interest.
We needed an AI agent that could call within minutes, sound natural, deal with interruptions and switches between Hindi and English, and pass the call to a human when the customer was ready to book.
How it works
The architecture, stage by stage. Click any stage to see why it exists.
The same pipeline serves outbound phone calls and in-browser conversations. Each channel has a transport adapter that normalises its audio stream, so the core agent never needs to know where the call came from.
Key engineering decisions
Channel-agnostic core
Telephony and WebRTC are thin adapters around one shared agent. Adding a new channel means writing a new adapter, not a new bot.
Latency as a budget
Every stage gets a slice of a sub-900ms end-to-end budget. Streaming STT, a fast inference provider and early TTS playback keep the conversation feeling natural.
Interruptions are expected
Barge-in handling stops the agent mid-sentence in under 200ms when the customer speaks, and turn management decides who holds the floor.
Benchmark, then commit
I compared speech models from OpenAI, Gemini, Anthropic, Deepgram, Google, Sarvam and Groq on cost per call and throughput before choosing the production stack.
Hard problems along the way
ChallengeSilence while the LLM or a tool call is working makes callers think the line has dropped.
FixContextual hold responses (short, natural fillers) play while tools run, so the conversation never goes quiet.
ChallengeCustomers switch between Hindi and English mid-sentence.
FixMid-call language switching lets the agent change its response language and TTS voice on the fly without restarting the call.
ChallengeSome customers want a human once they decide to book.
FixLLM-driven live call transfer hands the call to an agent along with the context gathered so far.
Impact
The platform now runs in production, re-engaging customers automatically and passing the most promising leads to the sales team.
- Booking conversions
- +10%
- End-to-end response latency
- <900ms
- Barge-in reaction time
- <200ms
- Pluggable TTS providers
- 4+