The educational and professional career path of Martin Vasilkovski is one that many data enthusiasts will find quite exciting. Being focused on academic research, complex mathematical problems, and challenging but innovative projects on one hand, while sharing a passion for sports and traveling on the other hand, talking with him on any subject is always a pleasure.
Our People & Culture
Being tremendously informed and insightful on the current global AI trends, explaining the most complex issues through simple enough examples that a 5-year-old could understand, this edition of our people & culture talk is one you cannot miss.
Education & beginnings in Data Masters
In the final year of my studies in Computer System Engineering, Automation, and Robotics, I developed a keen interest in Machine Learning (ML). My graduation thesis focused on creating a smart glove that could interpret and translate sign language and read it out loud, which I presented in an academic paper. This project marked my first encounter with ML, and after graduation, I immediately began working in the field. Although I enjoyed my work, I felt the need to further my education and enrolled in a master’s program in Milano, where I continued to pursue my passion for ML. Most of the courses I took were ML-related, and my master’s thesis involved developing a state-of-the-art solution for the da Vinci Surgery robot. Unlike a simple neural network, this solution required manual development, and I am proud of the work I accomplished. While studying, I also worked on applied ML at Siemens, but the position did not align with my personal goals and interests. During the COVID-19 pandemic, I became curious about the state of ML and data science in Macedonia and reached out to Martina Naumovska, a friend I had known for years. She encouraged me to contact her if I ever planned to return to Macedonia, and in January 2020, I joined the Data Masters team.
Machine Learning in Macedonia
It may be considered controversial, but I openly state that the ML education in Macedonia does not meet the required standards. To become a successful ML engineer, one must continuously expand their knowledge through online resources. During my master’s studies, I encountered eight different types of mathematics, which were essential for understanding the core principles of ML models. While it is possible to learn and perfect this knowledge, it requires a high level of focus and self-initiative.
In my opinion, the most suitable educational background for a prospective ML engineer would be FEIT, Faculty of Economics, FCSE, or Faculty of Physics, primarily due to the types of mathematics taught. Mathematics is crucial in this field. For instance, when working with data, it is imperative to analyze it, identify patterns, and observe its distribution. If the data is random, the model is ineffective. Therefore, understanding the data is essential for building a model. Critical thinking and analysis are indispensable skills for anyone working in the ML industry.
When the ML boom took place, Macedonian companies realized they lacked the necessary data. As global companies made significant progress, these companies realized they were falling behind. Currently, some companies want to implement ML and use models, but they cannot do so because they lack structured data. Many companies have attempted to use ML, but most have been unsuccessful. Although some companies use ML and develop models, there aren’t many of them.
Machine Learning in Data Masters & Current trends
At Data Masters, we have worked on projects across many domains. Our focus remains constant: data and analytics. We can provide solutions for companies that don’t have any analytics, develop Data Warehouses or cloud solutions based on their requirements, and support data-based decision-making processes using advanced statistics, algorithms, and ML. While deep learning and chatbots are useful tools, they are not always the best solution for clients. Often, simpler methods are more effective and efficient in addressing their needs.
Two years before Chat GPT, Data Masters worked and developed a model that is capable of understanding human communication, sublimating it into a few key phrases, and pointing out what is most important in the conversation. What’s interesting here: the technology and approach we had on this project are becoming popular once again, because ML professionals are starting to figure out the shortcomings of ChatGPT. I’m discussing transformer models such as BERT, for example. These models have undergone extensive training and have newer versions available. BERT, which was developed by Google, is considered to be the precursor to the current popular models. During our project, roBERTa was introduced, which is a smaller, faster, and more intelligent model that has since been optimized multiple times. Currently, this model has significantly improved and serves as the foundation for many advancements.
If you use GPT for a semantic search of documents, such as asking a question and receiving a response on which documents answer the question, the outcomes obtained through GPT and roBERTa differ by 10 to 15 percentage points, with roBERTa performing better. This demonstrates that while the current surge in machine learning is impressive, it is also quite consumer-driven.
When developing a final, detailed solution to address a specific issue, it is crucial to provide the best possible outcome. If I may say so, ChatGPT is average in everything, but not an expert in anything. In the realm of machine learning, achieving the highest level of accuracy requires precise predictions, which cannot be accomplished with an average model. This is why this scenario is very interesting and older models are being revisited, as it turns out that we jumped a few steps because we didn’t get the maximum potential from the previous models.
Advice for aspiring ML Engineers
When working with interns and young employees, I always emphasize that data is the foundation of everything. It’s impossible to create effective decision-making models without knowing how to work the data. 80% of the solution lies within the data itself. To accurately identify and solve problems, you need to understand the data and analyze it using statistical methods. Having a strong analytical mindset is crucial in this field.
Sometimes, only a small fraction of the available data is needed, even when dealing with vast amounts such as 10 terabytes. Overloading a model with unnecessary data can lead to increased personnel requirements, longer processing times, and higher costs. Therefore, it is crucial to understand all the underlying processes to gain a comprehensive understanding of how the model operates. Once you have a firm grasp of the data, you can determine which machine learning approach is most suitable for the task at hand.
Not every element is necessary for a model to succeed. For instance, if the model is highly dynamic and susceptible to noise, I eliminate neural networks from the outset since I am aware that they require smoothing out of information. Instead, I opt for decision trees or another type of model. I have spoken with many professionals and we agree that only about 15% of a machine learning project involves model creation. The remaining 80 to 85% of the project is devoted to data preparation, analysis, and labeling, which is akin to teaching a child to differentiate between good and bad examples. Despite popular belief, the majority of the work does not involve model building.
Personal (and company) goals
I would like to showcase the work we do for international clients in Macedonia. This will make our work more presentable and tangible in our local market. I believe it is important to demonstrate the capabilities we possess for both domestic and international clients. Serbia has a strong machine learning community, and we are working towards building a similar community here. We take it upon ourselves to raise awareness about our work and strive to achieve this goal. Additionally, I aim to use the knowledge we have gained to help in different applications, including business and scientific research.
The knowledge-sharing mindset between the employees is one of the most important components in the Data Masters growth and success in the past four years.
If you want to work on world-class projects with experienced professionals and you are not afraid to make mistakes (we learn and grow together), you are the perfect fit for us.
Click HERE to see the job openings!
Apply and master your talent with us.