About the Company: One of our well-established start-up clients in Palo Alto, CA are looking for an experienced Lead Generative AI Engineer/Scientist to train, optimize, scale, and deploy a variety of generative AI models such as large language models, voice/speech foundation models, vision and multi-modal foundation models using cutting-edge techniques and frameworks. In this hands-on role, you will architect and implement state of art neural architecture, robust training and inference infrastructure to efficiently take complex models with billions of parameters to production while optimizing for low latency, high throughput, and cost efficiency.
Responsibilities:
- Architect and refine foundation model infrastructure to support the deployment of optimized AI models with a focus on C/C++, CUDA, and kernel-level programming enhancements.
- Implement state-of-the-art optimization techniques, including quantization, distillation, sparsity, streaming, and caching, for model performance enhancements.
- Spearhead the development of Vision pipelines, ensuring scalable training and inference workflows of 10s and 100s of billions of parameter foundation models.
- Should be able to innovate for the state-of-the-art architectures involving Panoptic Segmentation, Image Classification and Image Generation. It is expected that the candidate experiments with the internals of Vision Transformers and convolutional Models like ConvNext, CLIP, Visual Question Answering (VQA) and Diffusion Models. Practice around AI Arts, Image Prompts, Conditional Image Generation will be an additional advantage.
- Design, develop, and innovate state-of-the-art in large multimodal models like GPT-4o, Gemini, Chameleon. Make architectural choices across dense / Mixture-of-experts, early fusion / deep fusion, choice of modality encoders (VQ-GAN, ViT, CLIP/SigLIP), decoders (Stable diffusion, Stable cascade, AudioLDM).
- Execute training and inference processes with a key emphasis on minimizing latency and maximizing throughput, utilizing GPU clusters and custom hardware.
- Innovate on current model deployment platforms, employing AWS, GCP, and GPU clusters, to enable high scalability and responsiveness.
- Integrate and tailor frameworks such as PyTorch, TensorFlow, DeepSpeed, Lightening, FSDP, and Habana for the advancement of super-fast model training and inference.
- Advance the deployment infrastructure with MLOps frameworks such as KubeFlow, MosaicML, Anyscale, Terraform, ensuring robust development and deployment cycles.
- Enhance post-deployment mechanisms with exhaustive testing, real-time monitoring, and sophisticated explainability and robustness checks.
- Drive continuous improvement initiatives for deployed models with automated pipelines for drift detection and performance degradation.
- Lead the charge in model management, encompassing version control, reproducibility, and lineage tracking.
- Cultivate a culture of high-performance computing and optimization within the AI/ML domain, propagating best practices and knowledge sharing.
Qualifications:
- Ph.D. with 5+ years or MS with 8+ years of experience in ML Engineering, Data Science, or related fields.
- Demonstrated expertise in high-performance computing with proficiency in Python, C/C++, CUDA, and kernel-level programming for AI applications.
- Extensive experience in the optimization of training and inference for large-scale AI models, including practical knowledge of quantization, distillation, and Vision Pipelines.
- It will be of additional benefit if the Candidate understands Diffusion Models (DDPM), Variational Autoencoders, Bayesian Modelling, Stochastic Variational Inference (SVI) and Reinforcement Learning.
- Experience in building 10s and 100s of billions of parameters generative AI foundation models.
- AI training job scheduling, orchestration, and management via SLURM and Kubeflow.
- Proven success in deploying optimized ML systems on a large scale, utilizing cloud infrastructures and GPU resources.
- In-depth understanding and hands-on experience with advanced model optimization frameworks such as DeepSpeed, FSDP, PyTorch, TensorFlow, and corresponding MLOps tools.
- Familiarity with contemporary MLOps frameworks like MosaicML, Anyscale, Terraform, and their application in production environments.
- Strong grasp of state-of-the-art ML infrastructures, deployment strategies, and optimization methodologies.
- An innovative problem-solver with strategic acumen and a collaborative mindset.
- Exceptional communication and team collaboration skills, with an ability to lead and inspire.
For more details, please reach out to Jia at jia@lagomtechnologies.com and setup sometime to discuss the role!
