Canada has built a reputation as a leader in artificial intelligence research. The more difficult challenge has been turning that expertise into viable companies.
Winston Li is trying to help close that gap.
His Toronto-based company Arima specializes in synthetic data: AI-generated datasets designed to replicate the patterns and behaviours of real populations without exposing their personal information. Its flagship product, the Synthetic Society, is a digital model of the Canadian population designed to help companies understand customers, measure campaigns and make decisions in a world where traditional data collection is becoming increasingly difficult.
“Even if someone’s your customer, you only know a snapshot about their life, and there are many more things happening you don’t know about,” says Li, who is also a part-time instructor at Northeastern University Toronto. A bank knows someone’s weekly transactions. A telecom company knows their phone usage. A retailer knows what size shirt they wear. But none of them sees the full picture. Information is highly fragmented, there are privacy concerns around sharing data and it’s time-consuming, expensive and difficult to conduct the kind of wide-reaching surveys necessary to gather this information. “You want to put all the information together, because when info starts to triangulate, you can learn new things.”
The Synthetic Society is a population-level synthetic dataset of 40 million Canadians made up of tens of thousands of attributes. Li explains it like this: if there are 100 people living in a certain postal code, Arima creates 100 synthetic profiles attached to that postal code with similar characteristics, patterns and behaviours. You won’t find an exact Winston Li replica at his own postal code, but you’d find a persona that comes pretty close.
The result is a dataset that companies can analyze without needing access to millions of people’s personal information. American research firm Gartner predicts that 75 per cent of businesses will use generative AI to create synthetic data by 2027, a massive jump compared to just five per cent of businesses in 2023. The Synthetic Society is already being used by companies like RBC, Rogers, Subway and Deloitte to get deeper insights into their existing and potential customers, offering a glimpse of what it looks like when Canada expands from inventing AI ideas to commercializing them.
Li grew up in midtown Toronto and didn’t stray far: he completed his undergrad in applied mathematics and graduate studies in statistics at the University of Toronto, which he describes as “just what all Asian kids did.” For Li, a first-generation Canadian, there was never a question that his life path wouldn’t involve studying “something like math, stats, engineering or computer science.”
When Li completed his PhD in 2013, data science was in its infancy, and there were very few people working in the field in Toronto, he says. “Back then, if you searched for data scientists on LinkedIn, maybe 300 people showed up. Today, it would be something like 50,000,” Li says. He was searching for a job in banking, finance, big pharma or biostatistics – the typical career paths for newly minted statistics PhDs. But when he was approached in his final year about working part-time at an ad agency, he was intrigued.
The agency, Omnicom Media Group, wanted to apply Li’s statistical know-how to marketing plans – a marriage of statistics with an intriguing sector he’d never considered. Working with companies like Scotiabank, Procter & Gamble and Honda, he spent his early career using statistics to determine things like whether sponsoring a sports team actually works as a marketing tool and where to place billboards targeting dog owners.
After moving into consulting at PwC, Li saw the same problem repeatedly: companies had plenty of data, but it was fragmented. He had learned about synthetic data during his academic work and began thinking about how it could solve that problem. Synthetic data had largely been used in fields like health care and academia, where researchers need realistic datasets without compromising privacy. Li’s bet was that the same technology could transform how businesses understand consumers. “We wouldn’t touch original data,” he says. “We would just make copies of datasets because at the end of the day, as data scientists, we don’t care that the first line of this dataset is John or Bob. We care about the signals that are in the dataset.”
In 2018, Li joined the development team for PyOD, one of the most widely used Python tools that scrapes data and identifies points that deviate significantly, indicating entry errors or even outlier events like credit card fraud. A couple of years later, he started Arima and began developing SynC (Synthetic Data Generation via Gaussian Copula), which removes those outliers using PyOD, then uses machine learning to merge datasets and scale them.
His original idea was to sell synthetic datasets to companies, but he quickly realized that customers needed more than data – they needed tools to use it. So he shifted to building the software. Reflecting on his time working in marketing and the most commonly asked questions he received, Li tweaked his methodology and renamed the program the Synthetic Society.
Today, Arima’s Synthetic Society contains tens of thousands of characteristics about consumers, from demographics and geography to purchasing behaviour, media consumption and lifestyle patterns. The company has built versions for markets including Canada, the U.S., the U.K. and Germany, drawing on census data, national research, market surveys, studies by private firms and credit card transactions. The company feeds this data to an algorithm that creates a synthetic set that replicates the relationships between points of the original one.
Li explains that this might involve filling in missing information for a big cohort based on the information from a smaller cohort. Or randomly sampling a dataset and then replacing and rearranging the information so it still has all the same relationships between variables (like how age and income might be linked). Some companies that work with Arima also provide their own data to get more specific insights for their uses. Arima then combines that information probabilistically with the national synthetic set and scales it to the target population. “Their data is very specific, and what we have is multi-vertical and very general. Piecing those two together can help them answer a lot of questions they wouldn’t otherwise be able to answer,” Li says.
In 2025, Arima worked with Coca-Cola Canada, combining synthetic data with the company’s own information to help identify stronger locations for vending machines. This meant digging into demographics beyond age and occupation, getting into the other brands that potential passersby shop, the routes they take, what media they consume and more. Coca-Cola reported that this resulted in an average 43 per cent increase in vending sales.
Get our weekly newsletter – the people, places, and ideas revealing where Canada is headed.
Despite such proof points, Li is careful not to position synthetic data as a substitute for reality. “It’s not a replacement for raw data,” he says. “It’s an augmentation of real data.” That distinction matters because synthetic data can also inherit problems from the information it is built on. If the original dataset contains biases or gaps, those can carry over into the synthetic version. To mitigate this risk, Li says Arima uses reputable studies with big enough sample sizes to be statistically significant, and constantly improves the dataset with new information.
His next goal is to move beyond marketing. Synthetic data, he believes, could help sectors ranging from real estate to government planning make decisions about where to build, invest and allocate resources.
“The Synthetic Society has so much potential and so many more applications,” he says. “Being a marketing tech company is not doing the technology full justice.”
