Ximilar: Visual AI that helps many businesses

Author: Klára Petrovičová for fi.muni.cz

To begin with, what motivated you to study computer science?

I was already interested in computers in high school – I learned how to code, I helped my father set up math books in LaTeX, but I wasn't exactly a computer "geek". I was deciding between studying mathematics and computer science, and at that time a coincidence intervened in my life, which in retrospect turned out to be very lucky: I slept badly on the entrance exam for Matfyz in Prague (laughs). I wrote them pretty far below my capabilities and didn't get in. In retrospect, I am convinced that studying at FI MU and living in Brno suited me much better than it would have at Matfyz in Prague. I fell in love with FI MU.

You went on to pursue a Ph.D., why did you choose it?

After my B.Sc. in 2002, I got a unique opportunity to spend six months at a university in the USA and be part of a group that was doing world-class research. There, I saw up close for the first time what research in computer science even means and how interesting and exciting it can be to come up with, for example, a new algorithm, implement it, test it and see that one has come up with a solution that solves some problem in the best way in the world.

That's why I decided to try Ph.D. studies, after completing my M.Sc. at FI MU. And then luck smiled on me again, because through my thesis supervisor I got to Prof. Pavel Zezula, who spent the whole nineties in universities in the West. Around 2004, there were not many people in the faculty who knew how to do research really well and at the same time could get people around them excited and passionate about their field the way Pavel could. In addition, he was able to get funding for Ph.D. students as well, so I was able to concentrate fully on my research work and writing papers.

What did you do during your Ph.D. studies at FI MU?

In Prof. Zezula's group, I was involved in similarity-based data mining. An obvious application is visual search in image data: If you have a large collection of images, this approach allows you to quickly find those images that are visually similar to a given example. This is exactly one of the services that Ximilar provides – in this area we are directly building on my work at the faculty.

Would you recommend Ph.D. studies to current students?

If a person enjoys thinking about complex problems, going really deep into it, and also has some self-discipline in their work, I would definitely recommend a Ph.D. degree. That said, you can try to study a Ph.D. without these qualities, hiding behind a tree and taking a scholarship – there are students like that too, but that's obviously not a serious approach.

What made you decide to leave research at FI MU and start a company?

I enjoyed my research work at FI MU very much, I was absorbed in it, I devoted all my time and concentration to it. However, over time, I found that I wanted to see the results of my research work in practice. There was a period of time when I was deciding and sitting on two chairs for a while, but eventually I made the final decision and left the faculty for the commercial sector.

What does your company Ximilar do?

We help companies around the world make better and automated use of their image data. We have both domain- and application-specific services that are ready-to-use, but also a platform that allows you to train your own models – we can call this platform Machine Learning as a Service. I dare say that our platform is unrivalled in the world in terms of its capabilities and the quality of its user interface.

One of our services is visual search in image data. In addition to pre-built similarity models that are suitable for e.g. product photos, we have a platform for training such models "on demand". The customer provides us with examples of image data from their domain and possibly information about what is considered visually similar, and our system trains a model specifically for the data and need. Ximilar also does image recognition, i.e., automatic categorization, image tagging, object detection in images and videos, and so on.

Did the theoretical knowledge from your research at FI help you?

Very much. In the beginning, when we formally started the company in 2017, and I was leaving the faculty, we directly built one of our services on my research work. Together with my colleagues at FI MU, we developed several indexing and similarity search algorithms. At Ximilar today, we use algorithms that are directly based on that research. Because we use machine learning methods, we have three Machine Learning experts on our team – I personally do not consider myself a real expert in machine learning.

What is your inspiration for the solutions your company offers?

We communicate a lot with our customers and potential customers, which are large corporations, smaller companies and startups, about their ideas. Startups in particular often come up with very interesting applications. We like to build solutions and services together with them that are one step further than other available solutions. These collaborations and discussions often result in ideas for completely new features and services, which we then incorporate into our portfolio.

How are your methods used in practice?

For example, Ximilar did several successful projects in the fields of medicine and biotechnology. Most of them involve automatic recognition of structures and objects in microscope or X-ray images. Within one of these projects, we have developed a system with a multinational corporation that can recognize bacteria and filaments in a fluid sample, which can be used to treat water to make it drinkable. The way it works is that in an area where drinking water is hard to get to, a person takes a sample, puts it in a small handheld microscope, attaches a phone and takes a picture, which is sent to our cloud. Our neural networks figure out what kind of pollution it is, so that person knows how to treat the water to make it drinkable.

Do you think artificial intelligence is the future for the medical field?

Good question. There are two views on the use of artificial intelligence in medicine. The first is the classic one, which says that only humans should be responsible for decisions concerning the health, and often the life or death, of another human being. Leaving this decision to artificial intelligence, into which we cannot see properly, is ethically and legally questionable, to say the least. Of course, these systems can be used as an assistant to help the doctor find something relevant in a large amount of data, but the final decision is up to the doctor.

But we saw also another approach, from one of our US partners, a dentist, who is developing an app with us to detect inflammation from root canal x-rays. He says that in the US, it often happens that when a dentist recommends to patients that they should have a procedure, patients tend not to believe them and think they are just trying to get money out of them. And, it is well known that dentistry in the US is ridiculously expensive. The moment there was an independent assessment, things might be better. A neural network can't get a bad night's sleep and evaluate an x-ray any differently than it did yesterday.

What has been your experience working with companies that use your technology?

Mostly very good. Whether they are large corporations or startups. More and more startups trust us enough to develop a service with us that is essential to their business. They are actually willing to become long-term dependent on our technology and services. I think they appreciate both the technical quality of our services and our flexibility, openness and willingness to collaborate on the best possible solution.

So there is very close communication and collaboration between you and the companies? Isn't that communication challenging at times?

It is. In the phase when we're building the service with them, we often have a shared Slack workspace where several people from both companies communicate daily. While such communication can be challenging, I've been lucky enough at Ximilar to surround myself with people who are comfortable with this way of working, enjoy it and are experts at the same time. They accept that it's a bit more pressure at times and that you have to answer to the customer at 8PM sometimes. And they are as happy as I am when the project is completed, the customer is satisfied in the long term and appreciates the content and form of our cooperation.

How many of you are in the company now?

There are over ten of us in the core Ximilar team, plus another ten or so people who only work with us on specific tasks. In the beginning there were three of us, then we added two more people as co-founders, and gradually we grew to current size. And that's all without any external investment. I must say that we are happy and quite proud that we were able to build a functioning company in this way.

What advice would you possibly give to someone who is now deciding to start a startup?

Have your idea validated by someone from outside – institutions like the South Moravian Innovation Centre (JIC) have consulting programs for startups, which helped us a lot in the first phase. Talking to someone who has seen hundreds of ideas and startups, who is interested in helping you but also asking the right questions can be invaluable. Secondly, I would recommend trying to find collaborators who are passionate about the idea and who you can trust. I personally wouldn't be afraid to give these key people a stake in the company, because without a good team, there is no good company. And third: think twice before letting an investor into the company. Of course, it's easier to build a company with investment at your back, but you get used to money without work and often have to approach investors again in a year or two. In any case, the investors should bring something else of value besides money: help with building the company, specific contracts or at least contacts to potential customers.

Is there anything that surprised you when you started the company?

In retrospect, I smile a bit at all the things I didn't know at the time or had a bit of naive ideas about. The subsequent process of learning and discovery often felt like "exploration by struggle", but I really enjoyed it. I had to learn a lot about working in a team, because in an academic environment, what you don't do alone often doesn't get done. Over the course of those few years, I had to learn to trust and delegate because otherwise the company couldn't grow.