Voice AI & Infrastructure Engineer

Why this role exists

We build AI that works where the internet doesn't — edge-first computer vision, voice, and robotics deployed across defence, industrial, and infrastructure environments. Our conversational voice agents run in production and are scaling toward thousands of concurrent calls, with new capabilities like outbound calling and real-time passenger/operator intelligence on the roadmap.

Today the voice stack and its infrastructure are largely owned by the founder. This role takes that off the founder's plate and makes it yours. You will be the person who owns whether our voice agents are fast, reliable, and ready to scale.

This is an end-to-end ownership role — you build the conversational agent and the infrastructure it runs on. If you want to own a real production voice product (not a fragment of one), this is that seat.

What you'll own

The agent

• Conversational quality across our voice agents: flow reliability, turn-taking, and barge-in/interruption handling.
• Diagnosing and eliminating the failure modes that plague production voice agents — name/word mispronunciation, hallucinated TTS artifacts, agent looping, and inconsistent flow execution — and putting monitoring in place so they stay fixed.
• Prompt and flow architecture on real-time LLM voice APIs, with the judgment to know when a problem is a prompt fix vs. an architecture fix.
• Hinglish-first conversation design that sounds natural to Indian users.

The infrastructure

• The real-time pipeline: VAD → STT → LLM → TTS, tuned to a sub-500ms latency budget.
• Telephony and outbound calling (SIP trunking, inbound/outbound call handling).
• Multi-tenant SaaS infrastructure on AWS (ECS Fargate, SQS, MongoDB) — deployment, scaling, and pre-warmed worker pools so a traffic spike doesn't show up in p99 latency.
• Observability: call analytics, latency tracing, and audit trails so we can debug a bad call instead of guessing.

Must-haves (2–4 yrs)

• Shipped at least one real-time voice agent to production — not a demo. You've felt the pain of latency, interruptions, and flaky turn-taking.
• Strong Python and comfortable owning backend services end-to-end.
• Hands-on with an LLM/voice API (Gemini Live, OpenAI Realtime, or a cascading STT→LLM→TTS stack) and real prompt/flow engineering.
• Practical AWS (containers, queues, deployment) — you can ship and run your own services.
• A debugging instinct for non-deterministic systems: you can isolate why an agent looped or hallucinated, not just notice that it did.

Nice-to-haves

• WebRTC, SIP/telephony, or LiveKit/Pipecat experience.
• MongoDB or event-pipeline / analytics work.
• Hinglish or multilingual conversation design.
• Playwright / scraping, or any multi-tenant SaaS scaling experience.

Stack you'll work with

Real-time LLM voice APIs · Python · AWS (ECS Fargate, SQS) · MongoDB · WebRTC/SIP (direction we're heading) · LiveKit/Pipecat (under evaluation)