Bytes of Brilliance: Sergiu Ibanescu, Tech Lead (.NET) and AI Champion

Introducing Sergiu Ibanescu a seasoned .NET developer with over 15 years of experience. As a hands-on Tech Lead at Maxcode, one of his goals is to explore new trends. Needless to say, that includes AI, which he has been working with for the past two years and continues to actively experiment with today.

What personally excites you about AI — and what makes you cautious?

I am very excited about the increased productivity that comes with using AI code assistance tools. I am also impressed by how easy it is to explore new technologies or to prototype solutions nowadays – especially compared to two years ago, when I first used Azure OpenAI to draft a prototype for a potential code customization tool.

It was fun, but the models available at that time (such as GPT-3.5 Turbo) didn’t seem powerful and capable of producing consistent results. Additionally, I am a fervent supporter of using AI for code comprehension and concept explanation. However, I believe we need to continue improving ourselves and not depend entirely on AI.

I often hear “code is dead” but, even if AI is becoming very good at writing code, I still think we humans need to remain in the loop and should never push code we do not understand into production systems.

What are the biggest risks for companies adopting AI too quickly or without a clear use case?

One of the main risks I see is the high rate of change in this field. You may need to rewrite or even rethink large parts of your AI-based workflows – whether because new techniques or models appear, new cloud options emerge, or even entirely new paradigms arise.

Therefore, you need to wait until an SDK or AI cloud offering has stabilized. Otherwise, you’ll simply spend more time and money on rewriting or repurposing your setup. To avoid the risk of adopting AI without a well-defined use case, it is essential to start with a clear problem statement and validate that AI is the right tool.

From a client perspective, when does and doesn’t AI actually add value?

From a business perspective, I believe that everyone wants to automate as much as possible. AI adds value when its non-deterministic nature is not a critical concern. There are several approaches to mitigate this – such as hybrid systems or human in the loop – and, in most cases, they make the tradeoffs acceptable.

However, you need to be prepared for some degree of unpredictability. If you cannot afford that, for example in critical or highly precise financial systems where every cent matters, then it is better to go with the traditional approach.

Maxcode has introduced GitHub Copilot Business as part of the development process. How are you using it in practice, and what difference does it make for your teams and clients?

We introduced this gradually to help developers focus more on understanding the business logic by delegating certain tasks to GitHub Copilot. Examples include writing unit tests, generating common pieces of code, or even implementing a basic functionality for which there are plenty of examples in the existing codebase.

Additionally, GitHub Copilot offers many models to choose from, so you could use Opus for research, drive a plan with Sonnet and – in the case of simpler, well-defined tasks – even implement it with Haiku, for instance. For our clients, this results in a faster development pace while maintaining the quality, because we are still in the loop all the time.

How is Maxcode approaching AI internally? What does ‘experimenting with AI’ look like in practice?

Earlier this year, we launched a new initiative called the AI Guild. Essentially, this is a group of enthusiasts – including me – who are passionate about AI and dedicate time to experimentation. We take it in turns to pick a topic, and then present our findings to the rest of the AI Guild, to a larger group or in some cases to the entire company. We are focusing on Microsoft Foundry Agent Service right now, but in the future we want to explore other options as well.

What have been the most surprising lessons learned from Maxcode’s AI experimentation so far?

Developing AI agents, or AI in general, is typically presented as being very straightforward – almost no code required. You just need a portal, and you are done in minutes.

However, even for things as simple as a chatbot agent, there are many considerations. You need to think about the context or the specific knowledge the agent should have and how to equip it, which tools the agent needs and how to connect to them, how to secure it, how to observe its behavior, how to evaluate the outcome, how to automate the agent deployment, how to optimize its costs, and so on. So I’ve encountered a lot of ‘How?’ questions that are not immediately apparent when you start.

Additionally, as already mentioned, due to the fast pace of change the ground beneath your feet is constantly shifting. Something that works today may need to be rewritten in a few months’ time because of new models or new SDKs.

Which common misconception about AI among today’s business leaders would you like to correct?

Everybody wants AI. Some businesses even brand things as ‘AI’ when they are not. I would recommend approaching it with a clear understanding of all its facets. One common misconception is that AI is a plug-and-play solution that delivers immediate ROI.

In reality, it requires thoughtful integration, ongoing maintenance, and a realistic understanding of costs. For example, it is widely believed we are living in the ‘Golden Age of AI’ from a cost perspective.

Almost nobody expects AI usage costs to decrease; nearly everyone sees the current pricing model as being the minimum. Forcing AI into certain use cases may result in long-term losses.