AI Engineer

Embedding VC•Palo Alto, CA

19d•Onsite

About The Position

LiveX AI is building realtime, interactive AI avatars: digital humans that see, listen, and respond with natural speech and expressive motion. The AI Engineer trains, fine-tunes, and serves the generative models behind these avatars across the full model lifecycle, from data curation and training diffusion, video, and multimodal models to low-latency inference for live streaming.

Requirements

At least 3 years of hands-on ML engineering experience, or strong research experience with production-quality code, and a record of shipping deep learning models.
Deep expertise in PyTorch and modern deep learning workflows, with the ability to read, reproduce, and extend research papers.
Strong background in one or more of diffusion models, video generation, generative AI, computer vision, multimodal (audio and vision) AI, speech synthesis, or neural rendering.
Solid Python engineering skills.
Experience with distributed training on multi-GPU or multi-node setups.
Proven realtime inference optimization experience with technologies such as TensorRT, ONNX, Triton Inference Server, CUDA, mixed precision, or streaming pipelines.
Comfort working in ambiguous, fast-changing environments, with a proactive and ownership-driven approach.
A fast-learning startup mentality and eagerness to push what interactive avatars can do.

Nice To Haves

Experience with talking-head, audio-driven animation, lip-sync, or digital human systems; 3D Gaussian Splatting, NeRF, face reenactment, or avatar synthesis.
Publications at top venues such as NeurIPS, CVPR, ICCV, SIGGRAPH, ICLR, or ICML, or notable open-source contributions.

Responsibilities

Train and fine-tune generative models for talking-head synthesis, audio-driven facial animation, and full-body avatar video, including diffusion models, DiT, 3D Gaussian Splatting, GAN-based models, and transformer-based video generators.
Design data pipelines for large-scale video, audio, and multimodal datasets, covering collection, cleaning, labeling, alignment, and augmentation.
Optimize models for realtime serving with sub-200ms end-to-end latency using techniques such as distillation, quantization, pruning, caching, streaming inference, and custom CUDA/Triton kernels where needed.
Build and maintain training, evaluation, and deployment infrastructure on GPU clusters using PyTorch, DeepSpeed, FSDP, Ray, and Kubernetes.
Define and track quality metrics, including lip-sync accuracy, identity preservation, expression naturalness, latency, and throughput, and drive measurable gains.
Work with backend, product, and research teams to ship avatar features and iterate on user feedback.
Keep up with research in generative video, multimodal AI, and neural rendering, and bring useful techniques into the stack.