Research Scientist - Efficient AI 高性能ai大模型研究科学家

Canva Canva · Enterprise · Beijing, China · Information Technology

Research Scientist focused on AI efficiency for generative models (design, video) in multimodal domains. The role involves optimizing AI from low-level kernels to high-level model distillation, architecture design, and inference systems. Emphasis on pushing boundaries for powerful yet computationally efficient models, with contributions to production features and community publications.

What you'd actually do

  1. Designing and advancing state-of-the-art generative AI models across multimodal domains
  2. Improving AI efficiency across the stack, including:
  3. Kernel and graph optimization
  4. Model compilation and systems optimization
  5. Model compression techniques (quantization, distillation)
  6. Efficient model architecture design
  7. Scalable inference systems
  8. Collaborating cross-functionally with research, engineering, and product teams to bring innovations into production
  9. Translating research breakthroughs into impactful features within Canva’s core products
  10. Contributing to the broader AI community through publications at top-tier conferences

Skills

Required

  • foundation models
  • generative AI (e.g., diffusion models, LLMs, VLMs)
  • AI efficiency
  • Low-level optimization (kernel/graph)
  • Model optimization (quantization, distillation, compression)
  • Efficient architecture design
  • inference systems
  • Python
  • PyTorch
  • Transformers
  • Diffusers
  • Megatron
  • DeepSpeed
  • cloud platforms
  • clear communicator
  • collaborate effectively across teams

Nice to have

  • meaningful open-source contributions

What the JD emphasized

  • publications in leading conferences (e.g., NeurIPS, ICML, ICLR, CVPR)

Other signals

  • advancing the future of AI
  • experimenting with cutting-edge techniques
  • improving models for real-world quality and performance
  • GPU-accelerated efficient AI computing
  • design generation
  • video generation
  • pushing the boundary of creating AI models that are not only powerful, bot also computationally efficient
  • tackles AI efficiency from low-level kernel optimization to high-level model distillations
  • novel model architecture exploration
  • creating efficient inference systems
  • foundation models
  • generative AI
  • multimodal domains
  • model compilation
  • model compression techniques
  • quantization
  • distillation
  • efficient model architecture design
  • scalable inference systems
  • research breakthroughs into impactful features
  • publications at top-tier conferences