Tech Insights from Innovative Minds: Lucian Cucoș, Deputy of AI

With a solid background as a .NET Developer, Lucian has built extensive analytical skills grounded on data-driven insights and advanced processing tools. For him, AI is a tool to rethink how we build software, how we work, and how we overcome challenges in all industries. 

What sets Lucian apart is his ability to think ahead. The potential of today's AI makes him curious about what this technology can do in the future, and you can see this in his desire to take things to the next level. "What's next? How can we use AI to make workflows smarter? How can we automate the repetitive stuff to focus on the big picture? What will we need tomorrow?" These thoughts drive his work, ensuring ASSIST stays ahead of the curve. 

One of Lucian's biggest contributions so far is his project idea - WiseMate, an AI-powered solution that integrates multiple data sources and uses an enterprise RAG system to find information within the company faster and easier. It's a great example of how he approaches challenges by creating solutions with a real and lasting impact. 

More than a simple programmer, Lucian is the kind of person colleagues turn to when they're stuck on a tricky technical issue. Besides giving answers, he helps people understand the why behind the solution. 

Because he's always open to answering questions, we asked him a few. Let's discover the world of AI from Lucian's perspective:

My expertise is in larger-scale projects, so I'm biased toward having a model that solves the needs of the product first and foremost and is out of the box. Techniques like reinforcement learning (RL) are fine, but in the current landscape, there are a ton of models to choose from, and unless you're part of an R&D project, I'd go with a validate & discard paradigm, keep it simple. Clearly define what ethical and unbiased means for your scenario. Limit your use cases to scenarios you can control via prompt engineering. Setup benchmarks to detect issues with the model down the line and switch to a different one as needed.

During my career, I've moved from embedded development to desktop, multi-threaded client servers, web servers under my desk, VMs in data centers, and cloud-native enterprise solutions. I've switched tech stacks, programming languages & paradigms. I've also had the role of DBA for quite a while. Every switch required a significant change of mindset. I was baffled when I discovered APIs, amazed when my team and I first used AI to understand and parse data inputs, and proud when I coordinated large-scale DB migrations & archiving using automated data pipelines. Mindset, attitude, and initiative are key. And whatever you do, don't lock yourself down into a single technology. You'll regret it later.

Imbalanced datasets are something you must expect and have been a documented issue since before ML, for example, the high/low cardinality issue of data set indexes. Before going on to specific techniques like downsampling, upweighting, and changing the batch size, cover the basics: "do you have access to the original dataset or only after a transformation pipeline? Has the data been cleansed? Is the data aggregated? If so, over what period of time, and what's the aggregator? What's your dataset size, and can you ask for a larger subset? Can you wait longer for the dataset to enlarge over time? How imbalanced is the dataset? Have you used it for training, and if so, are the results satisfactory?" Sometimes, the solution is simpler than you think and can save you a lot of work.

Keyword extraction and named entity recognition (NER) are definitely the more often used techniques that I've encountered. I'd have to say that sentiment analysis is the most interesting, but I find summarization to be the most useful. To give some examples, my team and I have used the suite of Language Azure services to build solutions that understand what you ask for and place an order for you (be it food or other topic-specific products that you can order online), using keywords extraction and NER. An interesting one was content moderation, which fired alerts linked to inappropriate content using sentiment analysis.

Many don't know this, but ML dates back to the 1950s with a computer program that learned checkers. In the '70s "The Hitchhiker's Guide to the Galaxy" had a super-computer called "Deep Thought" which calculated the answer to "life, the universe, and everything" in millions of years by learning from the entire Planet Earth from its conception till the present day. Once IBM's Deep Blue beat a chess world champion in the '90s, it was clear: ML is very good at learning from a dataset, identifying patterns, and predicting an outcome.

Gen AI is comparatively new in terms of timelines. We're still getting used to it and its various use cases outside its artistically inclined endeavors in audio, video, and product design. Yes, it can write a song; yes, it can write your blog post for you, but the real interesting stuff is what it can do if you use it in a software solution, in the software chain, basically as software that hasn't been written yet, or a dynamically self-writing software if you will.

At first, the AI industry was about imitating the human way of thinking. An LLM mimics what a human gathers throughout its lifetime. We're able to make decisions based on our experience or KB (knowledge base). There's still a lot of "why, how & where" to use AI in large systems, and it's generally allocated to analytics. But it can do so much more. There's already a lot of work being done with multi-agent systems and AI orchestration of software systems (Semantic Kernell), but only at small scale and specific scenarios. In large-scale enterprise systems, AI could replace a lot of tedious work associated with generating data pipelines, workflows, services & API orchestration, traffic management, etc. I'd like to see more consistent behavior in this area, which will encourage widespread usage.

I think the most important aspect of AI becoming widespread into implementation rather than just for on-the-side use or slow analytics is the advent of faster data access. Let me give you an example: when GPT3 came out, the only recommended way to have the model access the dataset was to turn NLP (or the request) into SQL and run that SQL on the dataset (TBH semantic search directly on the DB was faster and better). Now, especially in a cloud environment, resources can communicate and harmonize with each other so you can have a more direct connection between datasets and AI, effectively creating a data fabric. So, data integration is key, and being able to easily integrate a reasoning engine or AI-based SIEM in your ecosystem is even more so.

I had the fortune of having a math teacher (high school) who encouraged us to find multiple solutions to the same problem, including a software one (yes, I built software for math class). If you know that, for example, an algebraic problem can be solved using algebra, geometry, vectors, pen & paper, or software, your solution space expands. Using this problem-solving skill set, I'm able to define multiple directions for a project, explain things to clients in various ways, find innovative solutions to bottlenecks, and sketch out many variations on the same software solution till the team and I find the best one. You're never without options. Be the person on the team that says, "...wait, there's another way".

Security consciousness in our industry is at an all-time high, as far as I can see. No customer will accept an AI-based architecture without documentation on security concerns, disclaimers on the vendor's website & security audit docs. Questions like: "Where is the model hosted? Do I have control over it? Does it use PII data secretly for training?" People are aware of how easily data can leak, and we have to guarantee its integrity. Like in all aspects of security, education is key: what's the difference between a model (e.g., GPT-4), hosting of a model (e.g., Azure AI Service), a website offering a chatbot interface to a model (e.g., ChatGPT) or an API based service offering access to a model (e.g., OpenAI API). Understanding these will help you build a clear picture for your customers and remove the intrinsic stigma that AI brings with it.

The most important thing when you're in any tech stack is to understand the ecosystem and think at that macro level. It's never about this bit of code and that. It's about how it all fits together and helps the project thrive. For years I was involved (as a .NET Dev) in many MS CRM integration projects for a lot of happy customers. Although the tech stack I used and MS CRM itself have been superseded by other SaaS suits, I honed my data engineering skills during that time and have turned it into a passion. Data management, processing & integration are essential to my current role. Data exists before software & AI. The latter always depends on the former. If you take care of the data flow and model it appropriately, you'll be a good software engineer.

Another thing you should know about Lucian is that he's a natural mentor. As part of the ASSIST Tech Challenge, he has worked with tech students, helping them think critically and approach problems creatively. With a practical and engaging mentorship style, it's clear he genuinely enjoys helping others grow. He's also shared his expertise at Open Doors ASSIST, where he gave interactive tech presentations to students and professors. These sessions show his ability to break down complex ideas and make them accessible, and that's not always easy to do in tech.

Outside working hours, Lucian is actively involved in charity work, volunteering with the ASSIST Humanitarian Foundation to support children and community projects. One example is moderating the 2024 Christmas Charity Concert, an event that brought over 300 people together and raised funds for renovating the auditorium of the "Ciprian Porumbescu" Arts College in Suceava. This is proof that his inclination to help others extends beyond his regular job, as he is also ready to positively impact the community. 

There's one more thing that may surprise you. Many don't know that when he's not working with algorithms, snippets, or numbers, he's creating art. With a natural talent for painting and calligraphy, it seems like an unexpected contrast to his analytical mindset. In every piece of art he makes, he brings the same level of precision and thoughtfulness he does to software projects. In fact, he even hosted a calligraphy workshop for our team, revealing a whole different side of his creativity. Should we also mention that he plays guitar?

There would still be a lot more to say about Lucian, his curious nature, and his accomplishments. Thus, we want to leave a trace of mystery, so we will conclude that Lucian's story demonstrates that innovation isn't just about technology but the people behind it. It's all about how we work together, how we learn from each other, and how we push each other to grow. 

• Don't miss our previous edition featuring Diana Bărculescu, Head of QA. 

 
Stay tuned for more Tech Insights from Innovative Minds!