NeuroDrift is a US-based AI voice and enterprise software company. We build real-time voice AI agents that run live on real phone lines, at scale, for enterprise customers. Our platform (CallDash) carries production call traffic every day, so latency, telephony quirks and audio edge cases are our daily reality, not a research problem.

We're looking for a real-time voice engineer who has shipped voice into production and, more importantly, kept it working when it broke under load.

What makes this role different

Most "voice AI" work is gluing an API to a websocket. This isn't that. We run a self-hosted media stack: our own SFU, our own SIP gateway, and stateful agent workers on Kubernetes. When a call fails, the cause could be in the carrier trunk, the SFU, the agent process, the speech vendor, or the LLM behind a gateway, and those are five separate services with five separate logs. The job is as much diagnosis as construction.

What you'll do

•⁠ ⁠Build and operate real-time voice agents end to end: telephony, audio pipeline, STT/TTS, LLM integration, tool calling

•⁠ ⁠Own latency. Chase every millisecond from end-of-utterance to first audio, and know which component owns each one

•⁠ ⁠Integrate and harden SIP trunks against real carriers and contact-centre platforms, including transfers, header passthrough and call correlation

•⁠ ⁠Run stateful audio workers in production: capacity, autoscaling, graceful drain, and what happens to a live call during a deploy

•⁠ ⁠Build the observability that makes a bad call explainable after the fact, not just reproducible

•⁠ ⁠Tune turn-taking. Endpointing, VAD, barge-in and interruption behaviour are where calls feel human or don't

•⁠ ⁠Join client calls,



present your approach, and defend your technical decisions

What we need

•⁠ ⁠3+ years Python, including async (asyncio, FastAPI or similar)

•⁠ ⁠Real-time audio in production: WebRTC, SIP, RTP

•⁠ ⁠At least one SIP trunk integration you shipped and then debugged under load

•⁠ ⁠Production experience with a real-time voice agent framework - preferably LiveKit

•⁠ ⁠Kubernetes, specifically stateful workloads rather than stateless web services

•⁠ ⁠A latency mindset. You reach for end-of-utterance, time-to-first-token and time-to-first-audio before a user complains

•⁠ ⁠Comfort debugging from logs and metrics in an setting you can't attach a debugger to

Nice to have

•⁠ ⁠Self-hosted LiveKit (SFU, SIP gateway, Egress, Agents). This is our stack, so hands-on experience running it yourself is a significant advantage.

•⁠ ⁠Speech vendor tuning: Deepgram, Whisper, Azure Speech, keyword or keyterm biasing, telephony-band audio

•⁠ ⁠Contact-centre integration: Genesys, Five9, NICE, warm and cold transfer, SIP REFER, UUI header correlation

•⁠ ⁠Noise suppression and audio enhancement in a real-time path

•⁠ ⁠Prompt and small-model behaviour in voice contexts, where a dropped instruction becomes dead air

•⁠ ⁠Cost awareness: per-minute STT/TTS billing, and what a stuck call costs The setup

•⁠ ⁠Fully remote

•⁠ ⁠High intensity, 50 to 60 hours a week. Startup pace, not a 9-to-5

•⁠ ⁠Working hours primarily IST, with availability for US client meetings

⁠You'll thrive here if you like owning hard problems end to end, you're comfortable in front of clients, and "it works on my machine" isn't in your vocabulary

📌 Voice AI Engineer (India)
🏢 Neurodrift
📍 India