Practical Machine Learning: Lessons from Research to Application
Machine learning (ML) has become a defining technology across industries, but its practical implementation often raises challenges that go beyond model development. To better understand these challenges and opportunities, we spoke with Nemanja Čukarić, Head of AI at ZenHire. With a background in semiconductor physics and extensive experience applying ML in diverse fields, Nemanja shares his perspective on the importance of data quality, industry applications, and the role of generative AI.
Exploring Our Guest’s Motivation and Background
Could you share how your journey from physics evolved into your current focus on machine learning and artificial intelligence?
My background is in semiconductor physics, where I spent eight years teaching and running research at the University of Belgrade. In 2017, I decided to step out of academia and move towards industry. The transition was eased by the fact that machine learning relies heavily on mathematical tools I was already familiar with, such as linear algebra and statistics.
While I sometimes miss the slower pace and deeper exploration of academic research, I’ve come to enjoy the pragmatism and direct impact of industry work. Building solutions and seeing them in action is genuinely rewarding, even if it comes with tighter timelines and shifting priorities.
As Head of AI at ZenHire, what does your role involve and what are your main areas of focus?
At ZenHire, my work centres on building practical ML solutions to automate and improve elements of the hiring process. We’re a small team of five, so my role is both leadership and hands-on engineering – everyone contributes to both.
Our main projects include developing assessments, such as English proficiency tests, and CVDeepMatch, which matches candidates to jobs using CVs and job descriptions. The work is varied: some days involve debugging ML models, while others focus on organising team discussions or setting priorities. It’s a challenging but enjoyable environment, with the whole team committed to creating tools that make large-scale hiring more efficient for clients.
Challenges of Machine Learning Across Industries
You have applied machine learning in very different sectors, from recruitment to oil and gas. What common challenges do you see when it comes to data quality and usability across industries?
Across all sectors, I’ve found that data collection, labelling, and management are the most resource-intensive and challenging aspects of any ML project. Data is often siloed across systems, incomplete in important areas, and requires substantial collaboration with domain experts before it becomes ready for modelling.
The data exploration phase is critical and often underestimated, as it can reshape initial hypotheses and modelling approaches. At ZenHire, we faced a particularly complex challenge with English fluency scoring, where the subjective nature of language assessment required several iterations of scoring methods before reaching a production-ready solution.
This reinforced my view that successful ML depends more on understanding and preparing data than on advanced modelling techniques – a lesson that applies equally in recruitment or geological survey analysis.
The STELAR project is working on improving data discovery and usability in the agrifood data space, where information is often scattered and hard to connect. From your perspective, how relevant do you think this is for users dealing with complex datasets?
The STELAR project represents the type of foundational work that industry urgently needs but often overlooks. While attention is frequently directed towards advanced AI models, in practice, data preparation, discovery, and management are far more critical to success.
The Knowledge Lake Management System (KLMS) developed by STELAR tackles a fundamental challenge: transforming scattered, heterogeneous datasets into AI-ready, semantically interoperable resources. For users working with complex data – especially in domains like agrifood where information is fragmented across multiple sources – this kind of initiative is essential.
By emphasising FAIR data principles (Findable, Accessible, Interoperable, Reusable) and semantic enrichment, STELAR provides the infrastructure that makes advanced analytics possible, even if this work receives less recognition compared to model development.
Perspectives on Generative AI
Looking to the future, how do you see large language models and generative AI influencing industry and everyday applications in the coming years?
Generative AI represents a significant technological leap. In under five years, we’ve moved from early transformer models to highly sophisticated systems that seemed unimaginable not long ago. Compared with pre-deep learning NLP approaches, the progress has been remarkable.
However, I see a lot of misunderstanding around appropriate use cases. As the saying goes: when you have a hammer, not every problem is a nail. People criticising LLMs for not subtracting decimals correctly should instead use a calculator – that’s exactly what calculators are for.
Generative AI is still very young and will evolve quickly, but I’m especially interested in its adoption timeline within traditional industries. Safety, compliance, and intellectual property remain major barriers before it can be widely deployed. While the technology is already proving useful in writing and coding tasks, I’m cautious about predictions that AI will replace most white-collar jobs in just a few years.
The future of generative AI will likely be gradual, nuanced, and closely tied to integration with existing processes and human expertise, rather than full-scale replacement.
Conclusion
From semiconductor physics to leading AI in industry, Nemanja Čukarić highlights how practical machine learning relies on data quality, collaboration, and infrastructure as much as on models themselves. His insights underline the importance of projects like STELAR, which provide the essential foundations for AI in complex sectors such as agrifood.
Learn more about machine learning and generative AI on the STELAR Blog and follow us on LinkedIn for more updates.
