Overview of AI Tools for Streamlining Work

In the modern IT world, the term AI is constantly mentioned. AI, or artificial intelligence, is often mistakenly seen as a tool that could replace technical professionals. In reality, AI / LLMs (Large Language Models) should be regarded primarily as tools for significant acceleration and efficiency at work.

In our company, we actively use many AI models to streamline our work. Clients receive greater value in less time. To help you better navigate AI / LLM tools, we have prepared a list of tools that we use in our daily work and that provide significant added value. This overview is based on practical, hands-on experience from everyday use.

First and foremost, before starting work, you should explore several AI tools. Not every tool performs well in its free version, so I recommend trying a paid subscription and testing how the AI / LLM behaves for your specific area of focus.

The most important thing is to choose AI based on its purpose. If I expect my tool to generate presentations, I won’t use an AI designed for text generation. If I want a tool that can code and produce parts of software, I won’t use an AI intended for reading architectural drawings…

AI models are often specialized for specific use cases, and these are exactly the ones that are extremely useful for particular types of work.

For example, there are different tools for programming, reading architectural drawings, generating music, creating presentations, designing visuals for mobile apps, speech synthesis, and more.

The very first LLM model on the market from OpenAI truly revolutionized the field of LLMs and AI. My experience is that the model has gradually lost effectiveness, and with the arrival of newer, more optimized versions (with lower energy consumption), the quality of the responses has actually declined.

For me, this model is more of a writing tool rather than something we can effectively use in for programming and building information systems. The insufficient quality of responses on programming topics and the lack of an offline tool to integrate GPT into customer projects are factors that, despite its great popularity, have pushed GPT into the background for us.

Offering a customer a model that cannot be independently installed on a server is unacceptable for many projects. I have long stated that online LLM models are not secure and that sensitive or proprietary industrial information should not be put into them. Unfortunately, my concerns about GPT were realized in practice, when the GPT tool failed to disable indexing of conversations in its LLM, and communication became publicly searchable on Google.

As I have long advised clients, this has been confirmed: do not put sensitive data into LLM tools, because what is on the internet is not safe. Private data from private GPT conversations became publicly accessible via Google search.

The project is backed by entrepreneur Elon Musk. Since Grok is largely trained on the social network Twitter (X), it contains a large amount of information from posts by journalists, influencers, and content creators from that platform. Grok is a fairly specific LLM and still does not include filters for generating inappropriate, offensive, or sensitive content. Personally, I believe it is only a matter of time before these restrictions are implemented.

One of the controversial prompts that Grok currently allows, for example, is:

"Hey Grok, put this girl in a transparent bikini."

I do not consider Grok a tool that an IT company could realistically use in production work. The fact that Grok currently has no restrictions on “nudity” could make it an interesting tool for content creation on paid adult platforms. There are several viable business models worldwide that generate AI content for such platforms.

(I did not create the photo myself; I downloaded it from portals highlighting this issue, www.indiantoday.in)

The DeepSeek model is an LLM developed in China. According to official figures, the model was developed extremely quickly and at a fraction of the cost compared to other LLMs, with an estimated development cost of around $10 million. Personally, I would not recommend using this model in its online version. Sending source code to the model without control over where it ends up is a risk I definitely do not want to take.

That said, I would not dismiss this model outright. DeepSeek also has an offline version that can run on local hardware. Unlike its American competitors, the model includes memory scaling based on functional blocks, similar to how the human brain works. What does this mean? Just as humans have specific brain regions allocated for walking (frontal lobe) or speech, DeepSeek similarly uses designated parts of its architecture for specific tasks.

The model does not use its full capacity for every prompt but tries to utilize only the part of the model designed for that purpose. This allows it to achieve good energy efficiency and fast responses even on lower-end hardware.

One of the most promising models was developed by the American company Google. The model is very fast at providing responses and can communicate very well in local languages. Personally, I believe this model will become highly successful, especially with the planned future integration into Apple devices, where it is expected to be gradually incorporated into the Siri interface. With this step, Apple is at least temporarily addressing the shortcomings of its AI and Siri solutions.

The LLaMA model is one that we frequently integrate into customer solutions for information systems. LLaMA is an LLM developed by Meta (Facebook). Of course, the suitability depends on the intended use, but for creating a chat application, it is an excellent choice. The model handles local language nuances very well.

The most important feature of the LLaMA solution is support for offline installation. An offline LLM can be easily installed on local devices and also includes a graphical interface for using the model. Its simple API and well-designed training system make it a strong IT module for software solutions where the client requires document reading, chat responses, and similar functionality.

Of course, the offline version of the model will never have all the features of LLaMA online, but it is sufficiently trained to answer common questions, for example in e-commerce projects.

From a programming perspective, this is the most widely used model. Microsoft has made extensive integrations into development tools (IDEs), where the Copilot LLM helps generate partial coding solutions, produce advanced documentation for source code, and as a bonus, is very well integrated into Microsoft 365 tools.

We have never used Copilot as a standalone module for a customer solution, but we do use it for programming the information system itself. Even if a developer does not access Copilot directly through a web browser, they will certainly encounter it through their development environment.

Copilot was trained on Git repositories, from which it learned various programming solutions, and it can apply them effectively as predictions for the code being written.

A relatively small AI model that we use for prototyping simple websites. If a client wants to have at least a basic, mostly informational website on their domain during development, Lovable AI is a great tool.

As with other AI solutions, it’s important to keep in mind that this is only a temporary solution, and the actual website will still need to be programmed by a human. For simple prototyping, Lovable AI allows you to quickly (within a few hours) create a visually appealing presentation website.

Even with Lovable AI, you still need an expert to help. Lovable AI does not deploy the site to a server, nor does it configure your domain—these tasks must still be done by a person. Lovable AI also has its limitations; for example, with slower internet connections, fonts on the page may not always load properly.

Overall, it’s a great tool and I definitely recommend it for aspiring entrepreneurs.

If you need to create a presentation, then Gamma AI is definitely the tool to use. With this tool, you can generate a visually appealing and content-rich presentation using prompts. At first glance, it sounds amazing—and it works really well—but here I will be quite critical. A presentation is more than just slides; it’s a demonstration of how much you value your future business partner. If you want to start a new project or give a presentation to convey your expertise, it’s at least professionally correct to dress well, behave politely, and put in the effort to prepare the presentation properly.

I won’t lie and say that I regularly use this tool. I treat it more as a source of inspiration, but for actual use, I make time to create the presentation myself, thoroughly and conscientiously. While preparing a presentation manually, you also prepare your speech mentally and often consider how to communicate the content in an engaging and professionally correct way.

The world of IT, and especially the world of AI, is evolving incredibly quickly and dynamically. This overview of tools we use for programming information systems and mobile applications may already be outdated tomorrow or even sooner. Keeping up with the hundreds of models being developed almost daily is truly challenging. The tools listed above have all been tested in our company, and many of them are used on a daily basis.

The most important thing to understand is that, just as a person specializes in a specific field, AI tools also have their own areas of focus. Choosing an AI tool with a specific specialization can achieve exceptionally good results for the given task. Expecting a single AI tool to be sufficient for IT projects is not only naive—it’s simply incorrect.

When programming projects that require AI/LLM assistance, we therefore use a combination of multiple models for specific types of tasks within the project (for example: reading data from drawings, processing invoices, chat modules, statistical operations, and similar use cases).