About the team
The mission of Canva Research is to develop the AI technology that powers Canva's Creative Operating System, empowering everyone to design. We bridge cutting-edge research and applied methods with real-world challenges, building the technology that enables Canva, as an AI-native creative platform, solving today's problems and shaping tomorrow's.

Formed in June 2025, Canva Research has grown fast with teams across Sydney, London, Vienna, San Francisco and Beijing. We work across generative image, video, audio, agents, evaluations, on-device and foundation models, all with the same goal: turning research breakthroughs into features that make design more accessible, more capable and more fun for everyone who uses Canva.

About the role
Canva's generative models are judged by millions of people who'll never read a benchmark. They just know whether the design/output looks right. Turning that judgement into something measurable and something we can train against is one of the hardest problems in our research stack.

You'll build the systems that make that measurement possible: agentic evaluators, MLLM-as-judge pipelines, and the benchmarking infrastructure that tells us whether a model actually got better. This is a hands-on engineering role. You'll take evaluation research on reward modelling, preference learning, and inference-time alignment and turn it into practical, production-ready systems that give research teams a fast, reliable feedback loop.

What you'll do

• Design, build and optimise the infrastructure for MLLM-as-a-judge evaluation systems, giving research teams scalable, automated feedback

• Implement and experiment with inference-time alignment techniques, including prompt engineering, RAG and in-context learning, to directly improve model output quality

• Build and maintain benchmarking pipelines that compare internal and external models on design-centric tasks, watching for benchmark saturation and contamination that would quietly stop the numbers meaning anything

• Build infrastructure for reward modelling and preference learning that feeds post-training (RLHF, RLAIF) and inference-time selection

• Analyse evaluation data to identify model failure modes, and turn that into clear, actionable recommendations for the research team

• Run offline evaluation suites and online monitoring, with regression detection that catches quality drops before they reach users

• Collaborate with research scientists and ML engineers to integrate the agentic judge system into the model development lifecycle

• Translate the latest research in LLM evaluation and agentic AI into practical, production-ready engineering solutions

You're likely a match if you have

• A strong understanding of generative AI models (diffusion models, GANs, transformers, VLMs/MLLMs) and their architectures, deep enough to build evaluation systems that hold up

• Practical experience building or working with agentic AI systems, including MLLM-as-a-judge or similar automated evaluation approaches

• Experience with inference-time alignment techniques: prompt engineering, RAG, in-context learning

• Experience with reward modelling, preference learning, or alignment methods involving human feedback

• A solid understanding of machine learning fundamentals, hands-on experience with PyTorch, and the ability to optimise training and inference code for speed

• Experience managing or optimising large-scale distributed model training or evaluation runs across hundreds of GPUs

• Disciplined coding practices, comfortable with code review and pull requests

• Familiarity with evaluation libraries and frameworks

Nice to have

• Experience linking evaluation metrics to downstream product or business outcomes, and diagnosing why they diverge

• Knowledge of data visualisation tools to communicate findings effectively

• A background or interest in human-computer interaction, design principles, or AI ethics

Other stuff to know
We make hiring decisions based on your experience, skills and passion, as well as how you can enhance Canva and our culture. When you apply, please tell us the pronouns you use and any reasonable adjustments you may need during the interview process.

All interviews are conducted virtually