Senior Voice AI & GPU Engineer

Job Description: We are looking for a world-class Voice AI Engineer with deep NVIDIA GPU optimization experience to help us build and tune an ultra-low latency, conversational voice agent. This isn't a job for someone who just hooks up standard wrapper APIs. We need a specialist who understands audio streaming protocols, knows how to squeeze every millisecond out of a GPU pipeline, and can gracefully handle complex conversational problems like user interruptions (barge-ins) and awkward silences. If you have built end-to-end voice bots combining Speech-to-Text (STT) ➔ LLM ➔ Text-to-Speech (TTS) and deployed them on bare-metal or cloud GPUs, we want to talk to you. What You’ll Be Doing: End-to-End Voice Architecture: Architecting and deploying real-time voice agents using pipelines like Whisper/Deepgram (STT), open-source or API LLMs, and ElevenLabs/Cartesia/Kokoro (TTS). GPU Infrastructure & Performance: Setting up and optimizing models on NVIDIA hardware using frameworks like TensorRT-LLM, vLLM, or Triton to minimize Time-to-First-Token (TTFT). Streaming Audio Optimization: Handling WebRTC, LiveKit, or custom WebSocket pipelines to manage real-time audio bidirectional streams. Prompt Engineering & Conversational Tuning: Crafting bulletproof system prompts and logic engines to prevent hallucinations, reduce token waste, and keep the bot strictly on-script. Technical Stack We're Looking For: Languages: Python (Advanced) and C++ (for custom audio/inference bindings). GPU Architecture: Deep familiarity with NVIDIA CUDA ecosystem, quantization techniques (AWQ, GPTQ, GGUF), and vLLM orchestration. Audio Engineering: Real-time VAD (Voice Activity Detection), handling latency, packet loss, and user interruptions natively. ⚠️ HOW TO APPLY (Strict Screening Requirements): Generic, copy-pasted templates will be declined immediately. Please start your application with the phrase "Low Latency Audio" so we know you read this. To be considered, your proposal must answer these 4 questions regarding a project you have previously built: Show Us Your Work: Provide a brief overview of an advanced Voice AI project you’ve completed. Please attach/link a GitHub repo, a video demo, or a detailed technical breakdown of your code. The Blueprint: What exact models/engines did you choose for STT, LLM, and TTS in that project, and why? Prompting Strategy: How did you optimize your system prompts and structural logic to force the LLM to reply concisely (saving tokens and lowering latency) while keeping a natural verbal cadence? The Battle Scars: How did you solve the following three challenges in your build? Latency: What exact tricks did you use to drop the turnaround time to human-like levels? Pausing & Interruptions: How did you handle user barge-in so the bot stops speaking the moment the human cuts in? GPU Bottlenecks: What hardware/VRAM bottlenecks did you hit while streaming the audio pipeline, and how did you resolve them?