Vaakyo
AI voice agents for Indian businesses
A platform I am building for AI voice agents and telecallers that speak Hindi, English and regional languages, handling leads, COD order confirmations, bookings and reminders over real phone lines.
- Python
- FastAPI
- React
- TypeScript
- MongoDB
- Redis
- Celery
The problem
A lot of business in India still happens on the phone. Confirming a cash-on-delivery order, booking a clinic slot, following up on a lead. These calls are short and repetitive, and they have to happen in the customer's language.
I wanted an agent that could make and answer those calls on a real phone line, and sound natural doing it.
How it works
A call is a loop. The caller speaks, the agent hears it, decides what to say, and speaks back. If it needs something, like a free slot or an order status, it calls a tool in the middle of the conversation.
Everything around that loop is the platform: a console to build and test agents, a queue that places outbound calls, webhooks that report what happened, and an MCP server so the agents can be managed from an AI client.
What I learned
Latency is the product. On a call, a one second pause feels broken. So every step streams. The agent starts speaking before the full reply exists.
A live call lives in one place. The audio for a call is held by one server process. A request about that call, like "hang up", can land on a different one. So requests are passed along to whichever process is holding the call.
Anything that can wait goes to a queue. Dialling numbers, sending webhooks and sending email all run in background workers. The part that handles live audio does nothing else.
Test before a customer does. I built simulated callers that talk to an agent and a judge that scores the conversation. It catches bad prompts long before a real call.