AI agents are moving beyond simple automation into collaborative design workflows requiring fundamentally different approaches to user experience, model training, and infrastructure than traditional ML systems.

In this episode of The Data Engineering Show, host Benjamin Wagner sits down with Paul Tune, Staff Research Scientist at Canva, to explore how the design platform is building agentic workflows, managing multimodal data pipelines, and tackling the unique challenge of teaching machines to understand aesthetic taste alongside functional design.

If you enjoyed this episode, make sure to subscribe, rate, and review it on Apple Podcasts, Spotify, and YouTube Podcasts. Instructions on how to do this are here.

Paul is a Staff Research Scientist at Canva, bringing nine years of experience building machine learning systems that empower millions of users worldwide. With deep expertise in large language models, reinforcement learning, and generative AI applications, Paul leads Canva's post-training efforts on LLMs designed for agentic design workflows. In this episode, Paul shares insights into how modern ML teams balance competing priorities - from data efficiency and GPU optimization to evaluation frameworks for subjective tasks like design aesthetics. His work bridging the gap between casual users and professional designers offers valuable lessons for data engineers and ML practitioners looking to scale AI systems across diverse user bases and complex product surfaces.

"What Canva is is that online graphic design platform for you to be able to design and kinda have this whole end-to-end process from the designing, the brainstorming, and using all sorts of tools in order to create a graphic design." - Paul Tune

"The whole vision really is to empower the world to design, and what that entails is to then have this entire end-to-end experience of designing on the platform." - Paul Tune

"I think a lot of it has to do with matching intent, so even for yourself, if you're using cloud code, some folks go down the side of, I want to plan very specific things about my design." - Paul Tune

"Whereas for a more casual user, they probably do come in without really having an idea of what they actually do want in the first place, and I think having a few options to kind of show, okay, these are kind of like a few designs that you might like as part of that, so that sort of helps to then eventually narrow down the intent." - Paul Tune

"I think there's a lot of very strong momentum around generative tools right now, and as part of that, Canva is also experimenting with adding generative tools within the product." - Paul Tune

"I think one of the bigger trends this year is there's been quite a bit of buzz around agents in particular, and Canva is no different in that aspect—we are working towards agentic workflows." - Paul Tune

"I think for us, the biggest challenge is that every time we do a rollout by an RL algorithm where we do have a sample that needs to be scored and then some level of feedback goes back to the model to then update its weights, we have to heat up specific APIs within different services at Canva." - Paul Tune

"I think the change has definitely shifted from when we started work on machine learning, where you kind of source the data and then train, to really like how do you evaluate because large language models have so many capabilities." - Paul Tune

"I think I try to keep focus because I don't think it's very feasible for me to cover every paper out there, even though there are lots and lots of exciting things that happen every day." - Paul Tune

"I think what I'm particularly excited about is applying these sorts of models into domains that are beyond what is very strongly verifiable, like mathematics and coding, because progress outside these domains has been a bit slower, but I do see at least some progress over time." - Paul Tune