What I Have Learned About Generative AI – and How I Use It in Engineering
I have been designing camera systems, access control and integrated security systems for over twenty years. This profession teaches you to look sceptically at every new miracle solution: too often I have seen an impressive-sounding technology end up as an expensive box at the bottom of a rack.
Generative AI turned out differently. Over the past months I deliberately set aside time not only to use it, but to understand what happens under the hood. I completed several trainings at Gerilla Mentor Klub – AI Security in Practice, AI-Native Vector Databases, Azure OpenAI and ChatGPT, AI-Assisted Vibe Coding – and what follows comes from experience rather than enthusiasm.
What I understood about how it works
A large language model does not know things the way a database knows them. It breaks our words into pieces and calculates the most likely continuation based on what it has learned. There is no intent and no verification behind it.
That single sentence has three very practical consequences:
- It is confidently wrong. It cannot sense when it steps beyond its knowledge, so an invented standard number sounds exactly as convincing as the correct one. Every technical figure, standard reference and price has to be checked.
- It only sees what we show it. The context window is its desk. When I add the actual site layout, the device list and the requirements to a question, the answer is in a different league from what a half-sentence prompt returns.
- The quality of the question is the input. A precise request that sets a role and a frame produces a completely different result from the very same model. This is a learnable skill, and it is where I gained the most.
Where it genuinely speeds up my work
- Documentation and technical specifications. The structure and the first draft take minutes. The professional content and the responsibility remain mine, but I no longer start from a blank page.
- Writing quotations. The calculation is my job, yet it is a real help in phrasing things clearly and for the specific client.
- Data analysis. Going through a long event log or device list and spotting patterns is faster than when I pieced it all together by hand.
- Development. This very site is a good example: the custom CSS effects, the scroll indicator and the optimised videos were all built this way. I am not a web developer, yet I solved them myself – while genuinely understanding what I was doing.
Where I stop
Confidential client data or a plan describing the vulnerabilities of a specific site does not go into a public AI service. This is not over-caution but a professional baseline: what has been sent cannot be recalled. Where such work is needed, a closed, enterprise environment is the right answer.
And most importantly: engineering responsibility cannot be delegated. When it comes to a fire alarm interface, a door on an escape route or a camera angle that is sensitive from a privacy standpoint, the decision is mine and my name is on it.
What I take away from this
Generative AI does not replace the engineer. It replaces the blank page, the repetitive typing and the searching. Professional judgement, field experience and accountability stayed exactly where they were. But those who learn to use it well gain more time for what truly matters: making sure the system serves the client, and not the other way round.
I am happy to talk about this. If you are also wondering where this belongs in security engineering, feel free to get in touch.
