We are looking for a highly experienced AI Agent Engineer who has hands-on experience in designing, developing, deploying, and operating production-grade Voice AI Agents.
This role is specifically for candidates who have already built and successfully delivered AI Voice Agents and have worked on real-world voice automation systems—not candidates who have only experimented with ChatGPT, LLM APIs, basic voice bots, or demo-level implementations.
The ideal candidate should built a self-hosted, GPU-served voice agent — real-time STT → LLM (agentic flow) → TTS, with barge-in, turn detection, and telephony integration over our existing infrastructure. This is not a "wire together some APIs" role. You'll own the pipeline end to end: model serving, latency budgets, interruption handling, and the agentic logic that decides what the agent does mid-conversation, not just what it says.
If you've only built voice bots on top of a no-code platform or a single hosted API (Vapi, Retell, Bland, etc.) without touching what's underneath, this role will be a stretch, not a fit.
What you'll do
- Design and build the full voice agent pipeline: VAD, streaming STT, LLM-driven agentic dialogue flow, streaming TTS, and barge-in/interruption handling.
- Own end-to-end latency — turn detection tuning, streaming at every stage, GPU serving optimization (batching, quantization, concurrency). Integrate the pipeline with Telephony Infrastructure
- Design and implement the agentic flow logic — routing, tool-calling, escalation/handoff logic, multi-step reasoning within the conversation — and know when not to chain multiple agent calls inside a live turn because of latency cost.
- Deploy and serve models on self-hosted GPU infrastructure (vLLM, Triton, or equivalent) — not calling third-party inference APIs.
- Instrument and own the numbers: WER, per-stage latency percentiles (p50/p95/p99), concurrency handled, conversation success/failure rate.
Must-have — non-negotiable
- Has personally built voice agent that handled real user or customer calls in production, at real volume — not a hackathon demo, not a course project, not a prototype that never left staging.
- Hands-on experience with the full pipeline stack: VAD/turn detection, streaming STT, LLM-based dialogue orchestration, streaming TTS, and barge-in handling — and can speak to tradeoffs in each, not just name-drop the tools.
- Direct experience with at least one voice orchestration framework (Pipecat, LiveKit Agents, or equivalent) — and can describe a real limitation they hit with it, not just what the docs say it does.
- Experience serving LLM/STT/TTS models on GPU infrastructure directly (vLLM, TensorRT-LLM, Triton, or similar) — not exclusively consuming hosted model APIs.
- Can talk numbers: actual latency percentiles they've hit in production, concurrency they've scaled to, and at least one real incident — what broke, why, and how they fixed it.
Nice-to-have
- Multilingual/Indic language voice AI experience (Hindi or other Indian languages, code-switching).
- Experience with real-time noise suppression / echo cancellation (RNNoise, DeepFilterNet, or similar).
- Agentic frameworks beyond the voice stack — LangGraph, CrewAI, or comparable — for the reasoning/flow layer.
- Experience with semantic turn detection (beyond fixed-silence VAD endpointing).