We build a Voice AI agent for customer support that answers inbound calls and resolves common issues end to end. Speech-to-text feeds Claude, which handles intent, multi-turn dialogue and tool calls into helpdesk and order systems; replies go to text-to-speech in real time. Complex cases escalate to a human with a full summary.
Natural multi-turn conversation in the caller's own words, no menu tree to fight through.
Tool calls into helpdesk and order systems — look up an order, open a ticket, issue a refund.
Speech-to-text in, text-to-speech out, paced for low latency so callers don't talk over each other.
When it's the wrong system, a human gets the transcript, the intent and everything already done.
Support calls are long-running and stateful. A graph gives us explicit state, retries and a clean escalation edge, instead of hoping one big prompt chain holds together for ten minutes.
Async streaming for the audio path, contract-first Pydantic schemas on every tool, and one boring service to hand over when we move off it.
Software Development Engineer at McCain Foods India, building production AI platforms: multi-agent LangGraph workflows, RAG over enterprise documents, and FastAPI services. Previously founding engineer and ML lead at n0c tech. B.Tech Mathematics & Computing, DTU — 9.37 CGPA, 4th in department.
Incorporated June 2024, Gurugram, India. We build voice AI for customer support — the agent takes the "where is my order" calls so the humans don't have to.
We take on a small number of builds for support teams who are tired of callers pressing 1. If your queue is full of "let me check for you", that is exactly the call this agent should be taking.
We want this in the Claude Startups program for API credits and technical guidance on low-latency conversation, tool use, and cutting hallucinations in customer-facing answers.
apply@ / contact@ / hello@dhruvkejriwal.me