Job Title

Senior AI Voice Systems Engineer (WebRTC, GPU, Real-Time Voice, 1000+ Concurrent Calls)

Overview

We are building a self-hosted AI Voice Agent platform for recruitment.

System requirements:

• 1000+ concurrent AI conversations
• 3000+ outbound calls (Twilio/Telnyx)
• Real-time voice, low latency, high quality

You will review, fix, and scale our existing system.

This is a systems + real-time media engineering role. Not a prompt/API role.

Core Responsibilities

MOST OF THE CODING DONE BY CLAUDE CODE

• Audit current architecture and identify bottlenecks
• Redesign for low latency and high concurrency
• Build call orchestration and throttling system
• Optimize GPU usage across multiple nodes
• Ensure:
• STT/TTS/LLM runs only after human answers
• Stable performance under burst traffic
• Improve:
• Latency
• Voice quality
• Turn-taking and interruptions

Core Requirements (MANDATORY)

WebRTC

• Strong experience with:
• RTP, ICE, STUN/TURN
• Real-time audio streaming
• Experience working with:
• LiveKit
• Daily
• Must understand internal architecture, not just SDK usage

Telephony

• Twilio or Telnyx programmable voice
• SIP call flows and call state handling
• PSTN ↔ WebRTC bridging

Voice AI Stack

• VAD, AMD
• Turn detection (endpointing, barge-in, silence handling)
• Streaming STT/TTS pipelines
• Full-duplex conversational systems

Audio Processing

• Voice isolation and noise reduction (real-time)
• Experience comparable to:
• Krisp level quality
• Echo cancellation, gain control, PSTN-quality audio handling

Modeling & Inference

• Fine-tuning STT and/or TTS models (required)
• Experience running models locally (no API reliance)
• ONNX / ONNX Runtime (required)
• GPU inference optimization:
• CUDA
• batching vs streaming
• latency tuning

Concurrency & Systems

• Experience handling:
• 1000+ concurrent sessions
• WebSocket streaming at scale
• Strong understanding of:
• backpressure
• queue systems
• event-driven architecture

Programming

• Python (async, multiprocessing, GPU integration) OR
• Rust (preferred for performance systems)

Infrastructure

• Fully self-hosted (no paid STT/TTS APIs)
• GPU setup:
• 8× RTX 5090
• 8× H200

You must be comfortable working directly with GPU workloads.

Critical Capability

You must demonstrate:

• Triggering AI pipeline only after human answer detection
• GPU-aware call throttling
• Scaling to 1000+ concurrent conversations
• Maintaining low latency under load

Nice to Have

• FreeSWITCH / Asterisk
• Kafka / NATS / Redis streams
• SFU/MCU architecture experience

Deliverables

• Architecture review (issues + fixes)
• Scalable system design
• Optimized orchestration and GPU usage
• Production-ready implementation