Michael Levot
Keeping localization in focus

interview by cathy martin

Sydney, Australia
The University of New South Wales and New York University
My post-grad work was published in a journal alongside that of my intellectual heroes. I was also briefly the top-ranked player in the world at my favorite board game.
Paros, Greece, and Harlem in New York
Mixing cocktails, watching football, playing board games, and talking philosophy

Localization hasn’t always been top of mind at large tech companies with hundreds of competing priorities. But that’s changing, and Michael Levot — head of localization at Canva — has seen this shift up close. Not only that, he’s been a driving force behind recent internationalization efforts at the Australian graphic design software company. The Canva platform is now available in more than 100 languages, and the push is paying off: about 60% of users interact with the software in a language other than English (LOTE). Under Levot’s watch, Canva’s international monthly users have grown from 2 million to 138 million.

A linguistics nerd at heart, Levot developed an interest in AI at his first job in language services. He was an integral player in launching Canva’s machine-translation (MT) tool — fittingly called Translate — earlier this year, and he spearheaded the internationalization of the company’s new generative (GenAI) suite of tools called Magic Studio.

These days, he finds the challenge of integrating localization into all levels of the business most exciting. We asked Levot how his team manages to keep up with the localization needs of a rapidly expanding company, including the strategies and technologies he employs to scale localization alongside growth.

What first sparked your interest in languages? Was it education, travel, or something else?

I didn’t have much interest in languages until I got to university. In fact, even though my family is Greek, while growing up in Australia, I resisted learning the Greek language. Being Greek was a big part of my family’s identity, and I resented most aspects of diaspora culture until later in life. I regret that now, but it’s probably a common experience for second-generation kids. Because of that, I’m the only person in my immediate family who doesn’t speak Greek. So, it’s ironic that I ended up with a career in languages!

It was education that eventually sparked my interest in languages. In my first year at university, I actually got my timetable wrong and ended up in the wrong lecture: linguistics, instead of the psychology lecture I was meant to be in. But the class was interesting, and I liked that linguistics took a really messy subject and created some structure for thinking about it in a rigorous way. Language typologies, syntax, semantics — that stuff got me excited about languages. In a roundabout way, I think that’s helped my career in localization because I’ve always thought about how things apply to multiple languages, countries, and cultures.

You have master’s degrees in linguistics and AI ethics, which seems like a winning combination for work in the language industry. Did you always intend to pursue a career in language services?

No, not at all! It hasn’t been a linear path, and I don’t think I knew the language industry existed until I had a career in it. I was happily teaching linguistics at the University of New South Wales here in Sydney and looking forward to a long academic career. My first job in the industry, at Appen, was meant to be a summer job. But the time I had there was a very exciting, high-growth period in Appen’s history, and there was just an unreasonable number of very talented and good people, so I stuck around a bit longer.

At the same time, that job raised many questions about whether I was comfortable with the work we were doing: providing data for AI to government, military, and big tech. That wasn’t exactly on my list of things to contribute to, and it made me want to be more deliberate with my career choices. My wife, Lizzie, was finishing her Ph.D. then, so we had some flexibility. We packed up and moved to the United States so I could focus on studying ethics.

It turns out AI ethics is a saturated field, so I started shoehorning myself into niches alongside people who could have substantial impact. It was a hard decision when deciding whether to pursue a Ph.D. or come to Canva, but it was made a lot easier by the mission Canva has and the group of people I’d be working with. Canva’s founders talk a big game about positively impacting the world, and they deliver on it, so my itches were scratched. As a bonus, Lizzie and I could start to put down some roots in our native Sydney.

Can you tell us more about your experience and responsibilities at Appen?

I started as an intern at Appen. It was a steep learning curve; I had to learn how to do a VLOOKUP to pass the onboarding test. I had a strong background in linguistics but not in the areas relevant to speech technology, so even my strengths weren’t particularly relevant. My work at the time was mostly cleaning and standardizing speech transcriptions and annotations. Machine learning works most effectively when speech data is transcribed and annotated in a standard way so the patterns are easy for the algorithm to identify. The approach was usually to build processes that reduce the human effort needed to improve the standardization and quality of datasets. For example, we would find a non-standard spelling or grammatical form and create scripts or regular expressions (searches) that would find likely occurrences of that issue to be reviewed by a native speaker.

I worked on a mix of quantitative and technical projects, mostly for government and intelligence agencies, and some really huge-scale projects for big tech. Those were a different challenge, as you needed to be very organized to stay on top of things. And towards the end of my time there, I started managing a team working on a project that combined the huge scale and rigorous technical requirements. I loved it; it was paradise for someone with my interests.

Since joining Canva in 2018, how have the company’s localization efforts grown, and what has been the most exciting part of that growth process?

When I joined Canva, maybe half a dozen people focused on creating a better experience for LOTE users. At that point, the priority was really just (“just”) figuring out how to launch Canva’s core product in as many languages as possible. Skip ahead five years, and there are huge efforts across the business focused on international customers — content, growth, and marketing, among others. The localization teams are only a part of that effort. That was always the plan; my colleague and mentor, Rachel Carruthers, knew we needed a model that would scale with the growth the business was experiencing. The only way that was going to work was having a hub of localization specialists dropping into areas of the business, laying the seeds for an international focus, and then moving on to the next area.

I think that’s ended up being the most exciting part of working for Canva — the organizational aspect of it. That was one of the first and most striking differences I found moving from vendor side at Appen to client side with Canva. At an LSP, the craft of language services is at the heart of things, but in a software company, it’s one of hundreds of moving parts, and building trust and alignment is at least as important as the technical craft. Building trust and alignment is messy and inexact, and it wasn’t my forte coming into the job — not just doing it well but also being motivated to invest my time and energy into it. I live for it now, though. Seeing phrases my colleagues and I have coined show up across the business or being consulted on a plan two years after an initial discussion? Bliss. We’ve done our jobs.

That was hard when Canva was a small startup, but it’s a different ball game at our current size. How do you continue to have an impact in an organization that looks and works completely differently every six to 12 months? We’ve been learning from scratch how to do localization in a business of 200, 1,000, and 4,000 people for an audience of 4 million, 50 million, and 100 million monthly customers. At one scale, it was more or less possible to be across everything happening and create bespoke solutions for everything that came up. Today, we rely on automation and a talented team to stay on top of things.

There have been a few important milestones along the way. Some that stand out to me include hiring our first language managers, the first time we had native-speaker experts who could guide us on language-specific needs. That changed everything, and our approach actually moved from the general to the specific, against the pattern of other developments. Forming our internationalization engineering team had a similarly profound impact on things. Being able to address the needs of developers with automation and DevOps improvements removed a lot of friction in internationalization and localization.

Can you give us some background on Canva’s new MT tool, Translate? What was the motivation behind developing the tool, and what did it take to make it happen?

The motivation was that we didn’t like translating content created in our own product. I’m sure our customers felt that pain, too. In fact, a blog post created by a Canva customer called “The Nightmare of Translating a Canva Project” was the first result that came up when you Googled “translating in Canva.” Ouch.

We got the opportunity to do something about it when our CEO and head of product reviewed product pitches for an upcoming launch event, Canva Create. My colleague Renee is a strong localization ally, and she got me in the room for a two-minute pitch. If you can print a sweating emoji, please do that here! But we had a really clear idea of the problem to be solved, and the concept of how to approach it came together naturally. We have some amazing engineers in our internationalization team and a wonderful designer, Ben, who volunteered his time, and they just ran with it from the initial concept.

Translate has now been integrated into Canva’s GenAI suite Magic Studio, which creates written, visual, and multimedia content based on user prompts. What was it like working on this type of cutting-edge AI technology?

Having a front-row seat to GenAI as it was becoming mainstream was great. I think it’s a common experience for us localization professionals that our friends and family have no idea what we do, so to point to something happening in the news and say, “We’re working on that,” was refreshing! It was also a bit surreal to be conversing about something we in localization are so familiar with in such an unfamiliar context, and it was sometimes uncomfortable. We in localization have been grappling with AI for decades, and I think we deeply appreciate both the capabilities and limitations of AI. The new level of attention to this domain is both welcome and a challenge. We’ve needed to build on the momentum while at the same time capitalizing on the lessons localization professionals have learned from the earlier waves of AI in our industry.

You wrote on LinkedIn that more is on the way for Translate, including human translation. Are you ready to share any details about that?

There’s a lot more coming. Eri Sasaki is leading the product effort now, and some of the ideas and use cases she’s coming up with are really exciting. I don’t want to steal any thunder. Still, our main focuses are creating a human translation workflow — which I hope will have launched by the time this is in print — and improving the automation of design, such as resizing text to suit changes in character length and replacing fonts for language pairs with different scripts.

What is Canva’s localization team focusing on now? Any challenges to realizing those goals?

The localization team leads — Robyn, Celine, Susan, and Irene — and I have a few strategies that we’re honing, most of them variations on the industry’s longstanding issues. I don’t think anyone is satisfied with the conceptual tools we have for measuring the quality and impact of localization investments, so trying some novel approaches in that direction is a big part of our plans. It’s a Goldilocks problem. Anything that really gets to the heart of what the business is interested in moving will be too hard to correlate directly with specific localization efforts — you won’t tweak your translation memory leverage and see a measurable impact on acquisition. At the same time, metrics that we can move, like errors per thousand words, are too abstract to interest the larger company. Our teams have a lot of discussion around different ways of approaching the problem. We’re experimenting with customer satisfaction scores (CSAT) and combining granular measurements of conversion rates into something of global interest to map our interventions to business objectives. Obviously, we’re not the first to think about this and try to innovate, but I’m hoping we can make some progress.

I think this challenge has interesting parallels with other business functions. For example, we’ve seen SEO go from a fringe concern — often underinvested in and considered a “dark art” — to a core strategic function of successful businesses over the last 10 years. That has required change from both SEO specialists and the businesses they’re a part of; SEO specialists have had to develop sophisticated metrics and data tooling that directly measure the impact of content, while businesses have become more concerned with the more abstract metrics relevant to SEO (impressions, sessions, bounce rates, and more). Localization isn’t in quite the same place, and I think that’s reflected in the largely unresolved tension around how to measure its ROI.

Is there anything else you would like to add?

Yes! I have a challenge to pose to the MultiLingual audience. Localizing static and motion creative is too hard. You need control over the whole creative process to ensure assets will work in a traditional localization tool, or you end up duplicating much of the effort to reproduce an asset in multiple languages.

So, to the localization technology people reading this: I’ll race you to a viable product that takes an unlayered image or motion design and allows a user to replace any text without altering the design. It’s not a simple problem, but every other week, some new, miraculous technology is going viral. There must be a better solution.



Subscribe to stay updated between magazine issues.

MultiLingual Media LLC