How Canva Uses W&B Registry to Streamline ML Model Deployment

Canva has grown from a consumer design tool into one of the world’s most-used creative platforms, serving over 150 million monthly active users and routinely recognized among the most innovative companies in enterprise software. Behind the product, a team of more than 100 machine learning engineers works on everything from generative models and recommendation systems to personalization algorithms and search quality improvements. Managing that volume of ML work cleanly — from experiment to production — became one of the team’s core infrastructure challenges.

Before adopting the W&B Registry, the ML team at Canva had no clear architectural boundary between experimental training runs and production-ready models. Everything lived in the same artifact space, and distinguishing which model should be deployed required interpreting a complex combination of tags. For a team of over 100 engineers working across multiple model types, this created constant noise: it was easy to lose track of which artifact was production-ready and which was mid-experiment. The deployment process was slow and error-prone as a result.

ML Platform Team Lead Thibault Main de Boissiere and his team implemented Weights & Biases Registry as the linchpin of a restructured workflow. The Registry sits at the center of Canva’s ML stack, acting as the handoff point between experimentation and deployment. Experimental runs tracked in W&B stay in the experiment tracking layer; only production-ready models are promoted to the Registry, where aliases make it unambiguous which artifact is destined for deployment or A/B testing. The broader stack connects the Registry with Anyscale for on-demand notebooks and distributed training, Nix for dependency management, and Amazon ECS for production deployment.

The impact was immediate clarity. Engineers across the team — including product managers who also have W&B access — now see only production-ready models in the Registry. The tagging complexity that previously drove deployment decisions was replaced by a clean promotion model. As Thibault described it, the Registry “simplifies our lives in so many ways” by removing noise and storing all production-level information in one place. The team also praised the W&B UI overall, citing helpful system metrics and accessible access management as qualities that made the platform straightforward to operate at team scale.

Canva’s team sees the Registry as more than a solved problem — it’s a platform they plan to extend further. Thibault described ambitions for using the Registry to transform how ML deployment works across the organization, positioning it as a key lever in evolving from a reactive deployment process to a deliberate, governed production pipeline. For a company investing as heavily in AI as Canva, having that foundation in place matters as the model surface area keeps expanding.