Build for Speed: Lessons from Empowering GenAI Research at Canva
summary
Canva runs GenAI research with over 200 engineers, supporting 100+ models and about 2k training runs/month, aiming for “Demo in a Day” prototypes via Gradio and Weights & Biases. The platform uses AWS, Snowflake, PyTorch, Hugging Face, Ray and a homegrown Arnold CLI that builds containers locally and submits jobs so training starts in seconds; only ~20% of experiments launch, so speed, flexible GPU choice, reproducible dev environments and abstractions that avoid user migrations drive adoption.
AI-written summary · about
chapters (10)
- 0:00Accelerating GenAI research at Canva
- 1:48Researchers workflow and needs
- 3:54Key requirements for research projects
- 5:16Rapid prototype creation with Gradio
- 7:02Speeding up model development iteration
- 8:46Arnold CLI for fast training jobs
- 10:35Supporting researchers with hires
- 11:18Platform teams strategy for GenAI
- 13:46Adoption strategies for new features
- 14:26User feedback for platform tools
related talks
26mCanva's AI Leap: Scaling with Ray and Anyscale
47mLessons learned from using generative AI dev tools
53mHow to Build and Scale Ultra-Lean Gen AI Startups
31mThe role of CDEs in Generative AI
18mAccelerate generative AI development with integrated data platforms
44mKeynote: Tracking the Pulse of Generative AI
19mDeveloper Centric AI
17mGenerative AI: Keeping it real