tekBlueprint

I can’t say that I am an active user of AI and that I have automated all the processes around, but I have some experience working with LLM/LDM. Therefore, I would like to share my thoughts on this topic.

The main area of application was related to my daily tasks as an architect, so the assessment is mostly from this point of view. For someone else, everything may be completely different.

Presales

The presale stage and subsequent formation of a technical and commercial proposal most often requires the analysis of a large amount of information in a short time. Moreover, the quality and completeness of this information is not always as desired for the development of a high-quality offer. There are several cases when AI was useful, we are talking about the use of AI as a RAG (Retrieval-Augmented Generation).

First, AI saves in cases when a technical specification contains practically no useful information. It literally happened that the technical specification is fifty pages long with a description of everything, room temperature, chair ergonomics, etc., and among this is half a page with a small list of functional requirements. Here, AI helps to find the important parts related to functional requirements from the entire document without its proofreading and manual filtering of unnecessary information. In this case, general control of the issued response at the time of missed requirements is still needed, since the AI can hallucinate or miss something.

Here, the opposite situations occurred, when, when proofreading voluminous technical specifications, I missed some points (mentioned in passing in the document), but which the AI found. With a high intensity of the flow of incoming assessments, fatigue accumulates and you begin to miss little things.

Second, expand the description. For example, if the technical solution lists the technical stack that will be used, then it is not necessary to manually compose explanations for everything why this is necessary, but it is enough to indicate the names and approximate purpose and then ask the AI to describe in more detail. A banal saving of time on what was previously done manually.

For example, it was

• PostgreSQL
• RabbitMQ
• Nginx
• Kubernetes

Became

• PostgreSQL is an open-source distributed object-relational DBMS known for its reliability and big data capabilities.
• RabbitMQ is an open-source message broker that implements the AMQP (Advanced Message Queuing Protocol) for distributed systems and microservices architectures.
• Nginx is a high-performance web server and reverse proxy widely used for serving static content and load balancing across application servers.
• Kubernetes is an open-source platform for automating deployment, scaling, and management of containerized applications.

And that’s basically it. AI was not used for:

• Formation of the general concept, system architecture.
• Decomposition and evaluation of tasks.
• Determination of stages, necessary resources and cost of the project.
• Formation of a technical solution.

I’m not sure that this is even possible without preliminary training on high-quality datasets specifically for such tasks. And even then, it is far from certain that AI will be able to completely replace a person. The price of error is quite high to trust a nonsense generator.

Project work

Here everything is rather modest. The only thing where AI helped was tasks like summarization. For example, meeting minutes are not always well structured, information is duplicated, follow-up actions are not always clearly highlighted, etc. In general, AI could “comb” the minutes well, make a list of actions and responsible persons with dates.

I have not found any other application yet. Just like in presales, everything that entails a certain responsibility for proposed decisions and choices is done without the use of AI (even as RAG).

Development

I have heard more than once that AI is actively used by some developers to write code, but I have never had a positive experience with this approach. No matter how many times I have tried. Perhaps after two decades of programming, hand-to-hand combat is too unusual. For now, it is faster and easier for me to write it myself than to double-check and clean up what the AI has done. Although I would gladly give the code generation to the AI, I have not been interested in doing this for a long time.

BUT, there is still one thing that helps a lot - automatic filling of JavaDoc. This is something that the AI does clearly better than me. Yes, the comments may not always be detailed enough, but it is much better than nothing at all. Or forced documentation just because something needs to be written.

As a bonus, sometimes it can be used for simple unit tests. But most often I have to rewrite it.

Writing articles

It doesn’t work for me, or you need to know how to cook. The output is just a set of loosely related facts, phrases, and sentences pulled from the internet. Even if some text is generated, you will have to edit almost everything.

The only thing that helps now is to quickly get a short definition of something. For example, the question is “What is SIEM?”.

Security Information and Event Management is software that collects, analyzes, and manages security events from various sources in real time. The main functions of SIEM include centralized collection of logs and security data, anomaly detection, automation of threat response, and reporting on the state of the system.

But again, it is better to double-check, since hallucinations are possible.

2D / 3D graphic

There is one task that requires 2D / 3D graphics. I used text-to-image, image-to-image , ControlNet and other generations. In general, everything is pretty good with 2D now, through ComfyUI you can build complex processing pipelines. In general, it suits me. For example, the image in the header of this site was quickly generated through FLUX.1-dev.

However, with 3D in 2023-2025, everything is pretty sad, if we talk about open generation models. Low quality, terrible mesh topology, transparent materials are not supported. Perhaps for professionals this is not a problem to fix, but I am not a professional and not even an amateur.

Conclusion

For me, using AI as a RAG is yes, as a tool for creating any content entirely is no (with the exception of 2D).

It’s not that I’m against progress, I’ve often seen statements in the style of – “Here, machines have replaced horses, whoever was against this is outdated and resists progress”. But it seems to me that the comparison is not entirely correct, a comparison with a conditionally free drug would be more appropriate. Which does well, removing part of the complex thought process from a person, but at the cost of losing the skills of analysis, evaluation, and decision-making in the future. That is, at some point, it will not be a person who will be an AI user, but AI will use a person as an interface for broadcasting its nonsense to the world :)