Voice AI Engineer (Real-Time Speech)

About this role

We are looking for a Voice AI Engineer (Real-time speech) to join our team! Our client is an Azerbaijani telecommunications company and Azerbaijan's largest mobile network operator. The main products are: Fixed telephony, Mobile telephony, Internet services, Wireless broadband, and Value-added services. The primary goal is to accelerate the client’s Data & AI initiatives via a secure, hybrid cloud foundation on AWS while systematically modernizing the IT estate as part of the AWS MAP 2.0 program.

Key Project Objectives include: Cloud Foundation & Landing Zone: Deploy target hybrid network architectures, establishing a secure AWS Landing Zone Accelerator (LZA) and hybrid Data/AI platforms on AWS. Security, Compliance & Sovereignty: Operationalize on-premises data de-identification and Format Preserving Encryption (FPE) tokenization (achieving zero raw PII in the cloud), fully adhering to Azerbaijani Personal Data Law No.

998-IIIQ and Critical Information Infrastructure Rules (Resolution No. 229). AI Chatbot & Real-Time Voicebot Implementation: Develop and operationalize a flagship Customer Care Voicebot (STT → LLM → TTS pipeline) and Agentic Chatbot targeting Responsibilities: Design, build, and operationalize end-to-end real-time STT → LLM → TTS (Speech-to-Text / LLM / Text-to-Speech) voicebot pipelines on AWS, optimizing for streaming speech-to-text, first-token LLM generation, and first-audio TTS synthesis.

Deploy and maintain production customer-trained Whisper (Azerbaijani ASR) and Azerbaijani TTS models as low-latency real-time endpoints on Amazon SageMaker and specialized GPU node pools (NVIDIA A100/L40S). Implement and manage the Bedrock Proxy Gateway on EKS for multi-model routing, priority queuing via Redis Sorted Sets, cost caps, and high-availability serving targeting ~200 rps without API throttling. Integrate voicebot and chatbot decision engines with core enterprise telephony and CVM platforms, including Avaya (voice telephony), Genesys (digital chat/omnichannel), and Pelatro (CVM offer decisioning and uplift models).

Establish LLMOps & MLOps pipelines using Amazon SageMaker Pipelines and MLflow for experiment tracking, model versioning, prompt/agent registries, automated evaluation harnesses, and RAG knowledge base retrieval. Build call and chat transcription pipelines to ingest, transcribe, and extract real-time insights (churn risk, dissatisfaction, intent, lead signals) into downstream decision layers. Enforce data sovereignty and privacy controls by integrating on-premises Format Preserving Encryption (FPE) and tokenization wrappers into ML pipelines so zero raw PII enters AWS cloud environments.