This interview has been edited.
The Center for Data Innovation spoke with Karen Burns, co-founder and CEO of Fyma, an artificial intelligence (AI) company providing data analytics of video streams, based in London, New York and Tallinn. Burns discussed the power of AI to build on top of existing hardware, and why Estonia is a great place for start-up culture.
Ayesha Bhatti: What inspired you to start Fyma?
Karen Burns: Fyma was started by myself and my co-founder just over four years ago and the inspiration came from a client project that we did with our previous company for a city. The city was using really expensive sensors to understand vehicle types for a single lane of traffic. Those sensors are both expensive to install and maintain, so Fyma was tasked with trying to use existing cameras to see whether computer vision could provide the city with data on a specific area.
We spent some time building a minimum viable product (MVP), understanding and training models. We realised that with a single camera, we could do the work of 12 of those sensors, costing 4.5 times less than those original sensor solutions, over the course of the sensor’s lifetime. We realised that the world is full of cameras, just sitting there, mostly used after something happens and not really for any sort of meaningful operational data collection.
This sparked the idea of Fyma. We decided to launch it to start working on this very simple solution.
Bhatti: You touched on this briefly, but could you explain in more detail how Fyma works, and why it is so different to what has been done before?
Burns: What we have done is built a video analytics pipeline, so video goes in from one end and data comes out the other. What is really special is that it’s really simple, really quick to get set up without any additional hardware installation whatsoever. Other companies very often will either come and sell you a camera that does the computations on the camera, which is a big hardware cost and vendor lock-in, or they will sell you an edge device that can perform these computations on-device. In both instances, you still require a hardware installation. Fyma removes this entirely.
The other benefit of Fyma is that we are able to teach our platform to recognise any object in about a day. So let’s say tomorrow you’ve got golf carts coming to do maintenance work because they’ve been repurposed for your gardening team, we just need 10 images of that golf cart in order to train a model to detect them. So within a day our clients can start detecting a new object. We achieve this with clever synthetic data generation techniques and process optimisation.
Synthetic data generation is a great way to bypass a lot of the challenges of using real data. First it drastically reduces model training time because you bypass data collection, and second there is the ethical aspect of not massively scraping the Internet or existing cameras for images for datasets. We are also looking to use LLMs to interpret the output data, which empowers the insights we collect without putting loads of human resource into it.
And just to round off on how Fyma works, it uses virtual sensors so you can draw on the camera’s field of view to pick up information such as dwell time, occupancy, footfall, traffic etc. So where normally a physical sensor would be placed, with Fyma it is simply drawn on the screen.
Bhatti: What are some of the real-world problems Fyma is solving?
Burns: We first off usually start with just movement analytics. Some of our clients have no idea how many people are in their buildings, so if there is an incident, they would not know if it is 1,500 or 2,200 people affected in this particular building. So Fyma is really relevant because not all tenants share that data with their commercial landlord. It is also key data in retail, not just office spaces.
We then build up from there. In indoor spaces, we can monitor the performance of specific areas before and after an investment has been made, such as into either rebuilding a food court, or rebuilding a section of an entire shopping centre or mixed-use business district. We can understand traffic flows in real-time, dwell time and much more. Because Fyma acts like human sight, there are so many use cases that can be covered within the built environment.
Financially, we increase the return-on-investment on existing technology, there is a massive cost saving when not installing hundreds of “dumb” footfall sensors, but more importantly you can monetise the data by recreating your leasing strategies, or do completely different dynamic pricing based on footfall and occupancy. There are multiple use cases one can get out of it.
Bhatti: Have you encountered any obstacles and if so, how have you tried to overcome those?
Burns: Obviously there is GDPR which is the first thing that pops up in people’s mind. Are you tracking people? Are you looking at faces? Are you capturing gender identity in any way? Fyma has been built in a way that our AI has never seen a human face. The way it identifies a human is by observing “components,” so looking at just an arm, leg, torso, or hands, and that is enough for the model to understand whether an object is human, by seeing parts of the human body. We take a similar approach to how our model identifies other objects. It is just made up of different components.
We also don’t save any video footage either. We can limit video previews to staff, even the Fyma team itself, and we have actually been building our model together with the Estonian Data Protection Inspectorate as this is where our engineering team is based. We decided to go to the local data protection authorities straight away and they loved it. Usually people hide from them, but we went very early on and said can we make sure that we’re building this in compliance with the strictest interpretation of the regulation so that we can be compliant anywhere else that we go. We are also registered with the UK Information Commissioner’s Office. We often go to conferences as an example of how to collaborate with state and do ethical AI.
The other challenge is that most people don’t really understand AI. Often business talk only about ChatGPT, but it is really one of hundreds of tools that you can use to improve your business. But when ChatGPT hallucinates, trust drops in all AI solutions. The way the market is going, people learning about AI through mainstream media is not really aligned to making informed decisions about AI. There is a messiness there.
Ethics is really important, and I am for more regulation than less on this front. It is not catching up fast enough.
Bhatti: What is it about Estonia that makes it such a vibrant ecosystem for startups?
Burns: Until very recently, I was on the board of the Estonian Founders Association, which is basically a club for founders who have raised a specific amount of money and gone to market. The community is nearly 200 people strong now, and lots of new members were added last year. Estonia was locked away under Soviet occupation for quite a long time and because of that, everybody is used to just getting stuff done. If you didn’t have curtains, you made them. My mom made all my clothing and coming from scarcity to capitalism, people wanted to think big, make their wealth, and get things they never had as children. Everybody had the same trainers, towels, the same radios, the same everything. So it has been a big cultural thing to be aspirational.
Secondly, their smallness, I mean it’s 1.2 million people, really helps when you are trying to build communities. Many people know others, and the Prime Minister is probably a phone call or two away just because people are so closely linked with each other. That smallness is also good for thinking globally. Estonia is too small for anything so you have to think globally from day one. Companies launching on big markets like the UK or US never leave that home market. Estonians have to leave the home market immediately.
Finally, Estonia is very digitally forward. I can do everything online, send documents to my bank, submit taxes in minutes, verify suppliers. The ease of doing business is very different.