Artem Rozinskiy, Insight AI: `AI stopped being an `experiment``

In this role, Rozinskiy has brought a number of technological solutions to market, including a closed-loop RAG assistant for medical and legal compliance, the AI agent `Digital Tech Lead,` an AI model for Ingrad, and other solutions.

Artem Rozinskiy led the development of solutions for global corporations, including Hyosung, overseeing the full cycle from data collection and model development to pilot projects and production deployment.

𝓲

Фото: Artem Rozinskiy's personal archive

In the interview, Artem shares his hands-on experience in building ML products, managing teams, and turning prototypes into fully operational business solutions. We also discussed the state of the Russian IT talent market, approaches to team building, as well as projects in healthcare and construction that Artem considers pivotal in his career - including the development of a medical chatbot for a well-known German pharmaceutical company.

Artem, hello! You have built a portfolio of products - including AI agents and forecasting systems - in your role as CTO at InsightAI, as well as experience as an engineer at a well-known IT company. What was your very first project in this field?

My first ML project was far less `high-profile` than the AI agents or RAG assistants we’re building today at InsightAI. It was a fairly down-to-earth but highly practical experience - a system for forecasting operational metrics for an industrial client.

At the time, I was working as an engineer, developing a predictive analytics model. We forecast deviations in production processes based on telemetry and operational data. It was a classic machine learning system, and that experience later proved invaluable in my industrial-sector projects. That was when I truly realized that most of the success lies not in the model architecture, but in data collection, cleaning, and interpretation.

Today, more than half of the products my team and I at Insight AI build focus on automating labor, logistics, and production chains. And in these projects, the preparatory stage - laying the groundwork for high-quality data - is especially critical.

One of our most recent projects was a computer vision system for detecting defects on a production line. We deployed a CV system that analyzes video streams from the line, identifies defects and anomalies, records events, and generates structured `incident cards.` I particularly enjoy projects where the impact is immediately visible - after implementation, the defect rate at the factory decreased. Preparation played a key role as well: my team and I personally visited the facilities several times to oversee equipment installation and ensure proper deployment.

Your projects - such as the spare parts demand forecasting system - are used in marketing and planning. What was the primary goal you aimed to achieve with its development?

In demand forecasting, this becomes especially critical: when lead times increase from a few weeks to several months, a procurement mistake can cost tens of millions of euros.

My objective was not to `build a beautiful model,` but to create a tool that could be trusted in real operational workflows - a model that clearly indicates what to purchase and in what quantities, so the company doesn’t freeze capital in excess inventory while also avoiding lost sales.

Demand management is always about balancing stock shortages, logistics delays, and overstocking. We focused on ensuring that the model provided regular recalculations and delivered forecasts in a format convenient for procurement teams.

When the system was fully integrated into planning processes and began delivering measurable financial impact - eliminating over-purchasing and freeing up several million dollars in working capital, it became clear that the goal had been achieved.

Just a few years ago, it seemed that artificial intelligence in data analytics and business overall would not become a mass phenomenon anytime soon. What do you think became the main trigger for its rapid adoption?

I believe the key trigger was the convergence of two factors.

On the one hand, technological maturity: models became sufficiently stable, accessible, and understandable for real business implementation. On the other hand, businesses found themselves operating in an environment of constant uncertainty - sanctions, supply chain disruptions, rising costs, and workforce challenges.

Under these conditions, AI stopped being an `experiment` and became a tool for survival and optimization. Companies began adopting it not because it was trendy, but because without it, scaling and making fast decisions became nearly impossible.

Today, we aim to build models capable of making autonomous decisions and accelerating business processes. The future of automation lies in autonomous systems where AI predicts demand and adjusts supply chains in real time.

Our project for a major automotive retail company (where we built an end-to-end SKU-level demand forecasting system) demonstrates that the Russian market is ready for rapid transformation.

You have repeatedly launched innovative solutions for major companies, including Nakhodka and Hyosung, and collaborated with Ingrad, Refinish. What is distinctive about working with leading market players?

The main difference is the level of responsibility.

In large companies, you can’t afford to `play startup.` Every solution must be robust, explainable, and fully integrated into existing processes.

Such clients value not just the technology itself, but the ability to translate a complex idea into a working system that doesn’t disrupt the business but carefully strengthens it. In these projects, architecture, security, regulatory compliance, and the ability to communicate effectively with business stakeholders are especially important.

As someone who has progressed from an engineer to the CTO of the AI company InsightAI, why do you believe innovation is especially crucial right now, and how can it be implemented in a practical, real-world way rather than remaining just a `lab experiment`?

Innovation is crucial today because the status quo no longer works. Traditional processes can no longer cope with the workload, the speed of change, or the volume of data.

But innovation only makes sense when it is embedded in the real business environment. I always try to start not with the technology, but with the question: what exactly is causing pain for the company, and how can it be measured?

If a solution cannot be explained to the business in simple terms and its impact demonstrated, then it is a laboratory experiment, not a real product.

You hold the role of a technical leader — managing a team, designing solution architectures, and working with the product. What makes it challenging to be not only an engineer but also a manager?

The challenge lies in constantly being between two worlds. On one hand, there’s technology, where you want to do everything correctly and elegantly. On the other hand, there are people, deadlines, business constraints, and responsibility for results.

As an engineer, you think about quality; as a manager, you think about balance. You need to make decisions that may not be technically perfect, but are optimal for the product and the team. All while maintaining trust from both developers and the business.

Let’s talk about innovation in healthcare. You have repeatedly implemented standout projects in the market, including a medical Q&A chatbot for a well-known pharmaceutical company. What was the main outcome of this project?

The main outcome was that we demonstrated generative AI can be used safely in medical and legal environments if the architecture is designed correctly.

For this client, it was important not just to have a chatbot, but to create a tool that relies solely on verified knowledge and avoids arbitrary interpretations. The result was a medical assistant that reduces the workload on specialists and standardizes responses, all while remaining fully compliant with regulatory requirements.

For me, this was an important case because it clearly showed how complex AI technologies can operate effectively in sensitive and highly regulated domains.

The so-called `vibe-coding` trend is becoming popular. How do you feel about this trend?

Cautiously. On one hand, it emphasizes an important point: engineers truly perform better when they aren’t afraid to experiment, can prototype quickly, and don’t get bogged down in perfectionism.

But as an engineering approach, it is unsound. In production, especially in ML, system software, or infrastructure - vibe-coding can lead to chaos. It’s particularly problematic when startups use it to create a quick MVP and then market it as a finished product, resulting in projects where no one fully understands the code.

I personally adhere to strict standards in ML engineering, since I work with industries like pharmaceuticals and other critical sectors where weak architecture is unacceptable. Solutions must be justified, free of technical debt, and - most importantly - scalable. I specialize in solutions designed for a long lifecycle.

Good teams know how to combine the freedom to explore ideas at the start with engineering discipline during implementation.

Author: Dmitry Kaminskiy