Senior Voice AI Engineer

Keka Technologies Private Limited; Bengaluru; On-site; INR 4,000,000 - 7,000,000; Full time

Job summary

Keka Technologies Private Limited in Bengaluru seeks a Mid–Senior voice AI engineer to own distributed, real-time systems that route thousands of concurrent calls at scale. You will design how STT/ASR, LLM, TTS, RAG, and tool-calling cooperate to keep latency low and quality high.

You will own end-to-end latency, diagnose issues, build measurable quality signals, and mentor a growing team while extending the platform toward live assist and agent support.

Qualifications

• 5+ years building production products and systems, with at least 2 years in real-time voice in production
• Deep understanding of the full voice pipeline and its impact on quality and latency
• Strong in Go and Python; production-grade, performance-critical code
• Track record of running voice AI at scale with high concurrency
• Fluency in provider tradeoffs between STT/LLM/TTS/RAG/tool-calling
• Eval-driven approach to quality with signals and real-time evaluation
• Comfort with agentic coding tools and treating quality as a measurable signal

Responsibilities

• Own voice AI at scale and keep thousands of concurrent calls fast by optimizing hot paths in Go/Python
• Own end-to-end latency, targeting sub-second response times across STT, LLM, TTS, RAG, and orchestration
• Diagnose and fix call quality issues quickly by tracing pipeline signals
• Build and run voice-quality evals to gate releases and detect drift
• Make platform tradeoffs between quality and latency for each conversation
• Extend the platform into live assist and agent assist features on the same pipeline
• Set engineering standards, review, and mentor a growing team

Skills

Voice AI at scale

Real-time concurrency

Go/Python

Provider tradeoffs fluency

Eval-driven quality

Agentic coding tools

Pipecat

Real-time voice framework

Tools

Pipecat

Real-time voice framework

Job description

Level: Mid–Senior (5+ years | 2+ years building voice AI in production)

About the Role

You will own the distributed, real-time systems that orchestrate parallel voice calls at massive scale: the core platform behind every conversation Convogent runs. The job is keeping thousands of simultaneous calls fast, reliable, and natural, owning how STT/ASR, LLM, TTS, RAG, and tool-calling work together, and where every millisecond and every point of quality goes across the pipeline.

What you’ll do

• Own voice AI at scale. Keep the pipeline fast and reliable at enterprise concurrency, thousands of simultaneous calls, by lifting calls-per-vCPU through extending Pipecat or rewriting hot paths in Go.
• Own end-to-end latency. Keep the conversation under ~800ms by decomposing the budget across STT, LLM, TTS, RAG, tool-calling, turn-detection, and network hops.
• Diagnose the bad call. When a call does not feel right, locate the fault in the pipeline fast by asking the right questions, reading the right signals, and owning the fix.
• Make quality measurable. Build voice-to-voice evals that gate releases and catch STT/LLM/TTS provider drift before customers do.
• Own provider tradeoffs. Decide how STT / LLM / TTS / RAG / tool-calling choices balance voice quality against latency, per conversation flow.
• Build adjacent products. Extend the platform into live assist, agent assist, and what comes next, on the same pipeline foundations.
• Set the bar. Review, mentor, and define engineering standards for a growing team.

What we’re looking for

• 5+ years building production products and systems, with at least 2 of those years in real-time voice, operating it in production, not just shipping a demo.
• Deep understanding of the full voice pipeline, every stage end to end and how each one affects the next and shapes quality and latency.
• Strong in Go/Python, comfortable across the whole pipeline. Production‑grade, performance‑critical Go is especially valued: you have rewritten hot‑path components in Go to lift throughput under real load.
• A track record of running voice AI at scale, scaling Pipecat (or a comparable real‑time voice framework) to high concurrency in production.
• Provider‑tradeoff fluency: how STT / LLM / TTS / RAG / tool‑calling choices play off each other on quality and latency.
• An eval‑driven approach to quality. You measure voice quality with real signals, not vibes, and have built or run evals on a real‑time voice or LLM system.
• Fluency with agentic coding tools and a habit of treating quality as a measured signal.

What we’re NOT looking for

• A prompt engineer / "AI app" builder who has only called LLM APIs and never operated a real‑time pipeline under load.
• Someone who has only used managed platforms (Vapi/Retell/Bland) and never scaled the layer beneath them.
• A pure web-backend CRUD engineer with no real‑time/streaming/voice exposure.
• An ML researcher who wants to train/fine‑tune models. We integrate providers; we don't build models.