Key Responsibilities

Own end-to-end architecture: ASR → NLU/LLM → dialogue manager → TTS → telephony

Make build-vs-buy calls on ASR/TTS vendors (Deepgram, Azure Speech, ElevenLabs, or open-source Whisper/Coqui)

Own latency budget across the pipeline (target sub-1.5s round-trip for live calls)

Mentor Conversational AI and ML/Voice engineers; own technical hiring bar

Must-Have Technical Qualifications

Production experience with at least one ASR engine (Whisper, Deepgram, Google STT, Azure Speech) at scale

Hands-on experience with LLM-based dialogue systems (function calling, RAG, or fine-tuning)

Strong grasp of telephony integration: SIP, WebRTC, Twilio/Exotel/Ozonetel or similar

Experience optimizing for latency and cost simultaneously in a live voice pipeline

Comfortable working across Hindi/Hinglish and English accent handling for Indian/UAE markets

Track record of hitting production accuracy benchmarks: 95%+ transcription (WER-based) accuracy and 90%+ intent classification accuracy at scale — not just in a demo





Experience handling dialect variation and code-switching (e.g. Hinglish, Gulf Arabic variants) in production ASR/NLU

Valuable-to-Have

Prior experience at Amazon Alexa, Microsoft Cortana/Speech, Google Assistant, or a voice-AI startup

Exposure to Arabic ASR/TTS for UAE market

2+ years hands-on with LLMs/Generative AI in a production (not POC) setting

Screening Questions

Describe a voice pipeline you've architected — where was the latency bottleneck and how did you fix it?

How would you handle Hinglish code-switching in ASR for an Indian BPO client?

Whisper vs a commercial ASR API — how do you decide, and what's the cost/accuracy tradeoff?

How do you evaluate hallucination risk in an LLM-driven dialogue manager for a live customer call?

Walk us through how you measured and hit a 95%+ transcription accuracy target in a past role — what was the failure mode you had to fix to get there?

📌 Lead / Voice AI Architect (Telangana)
🏢 Technotask
📍 Telangana