four years of AI in my engineering life
My AI journey until today
The topic I want to share is one that had the most influence in my business and engineering life over the last few years and it does also have the fastest transition at the moment.
I want to start giving you a short introduction about AI models and how I experienced the last 4 (and it's crazy that is only 4!!) years with it.
At the start: This post is for people that are somehow known to software development or just curious about the topic. I am no classic software engineer with an computer science background. I am more the hardware engineering guy that somehow drifted into the software side of things and now software (for hardware) is more of my daily business than hardware.
ChatGPT: The beginning in 2022
If you have tried out the first model of ChatGPT I think everybody had this "wow" moment. Sure the answers were still a bit hallucinated and the first use cases me and my fellow students tried out were simple calculations and let him write some funny stories with us in the main role.
Back then I was part of the electric subteam of the formula student team at my university. If you are not familiar, this is basically a bunch of engineering students that try to build a fast race car and quickly find out that most of engineering is not about building and more about debugging and fixing things.
I was responsible for the whole electric system there and as we had ~10 custom PCB's located in our car with an microcontroller on it we needed to write a lot of software. This was the first time AI was generating code for me - it was still pretty simple: Copying code that does not work into chatgpt app => trying to give him as much information about the rest of the system => copying the answer back into the application and hope the error is fixed. Compared to the workflow now, this was super stupid but it was the start.
GitHub Copilot: AI enters my IDE in 2023
In 2023 Copilot integrated codex and for the first time you had an agent that was sitting inside your vscode window and you could handover specific files he should include in its reasoning and he could directly edit the code in the file.
This felt like magic back then - generating small python scripts and tools were a no brainer, CAN databases with more than 1000 lines of code could be generated and verified in minutes than going through them by hand, writing application code was starting to go more into describing the problem and approving solutions than doing all the handwork. But it was still very guided work: You needed to know what file you want to edit, how the architecture should look like and where to search for bugs.
Claude Code: The first real coding agent in 2024
In 2024 the first terminal coding agents arrived. Back then I did not realize them too much, because they were token based at first. I think claude later then released a 20$ plan and I tried it out.
And wow: This was a whole other level than anything copilot was capable of. Now claude could check your whole working directory and make edits to all kind of files and this being much more capable than copilot ever was (in my experience). You started creating todo lists for claude and he would work them off in a 30min run, you are doing reviews of the work and you got a 300% speed up.
To guide your agent you would then create .md files with many information and rules about the environment and what standards you are having in your code base, what is your coding style and so on. So the code got more and more "human" and fitted your earlier code (..and was even better).
I improved and refined this setup more and more. By 2026 Claude Code had basically become a normal part of my development workflow. I wasn't really looking for another tool anymore because my setup worked well and I had spent a lot of time refining it.
Codex Astra: AI enters my hardware workflow in 2026
Then in September OpenAI has released Codex Astra model. To be honest: I did not take all these ai model benchmark tests too serious and tried to get the best working setup. My claude code setup was pretty good for my use so I just stick to it and improved it rather than adapting to a new model. But the astra release showed a PCB routing with KiCad and here I got curious: An AI agent that could help an engineer during the routing process sounded to good to be true - so I had to try it out.
Until this point, AI had mainly changed the software side of my engineering work. The Astra release was the first time I really started thinking about what the same development could mean for my hardware work.
I took a small PCB I had designed some time ago that was rather simple, maybe 50 components placed on it. I did two tests:
- Super dumb just prompt him - here is the schematic, please finish. You can guess - the result was more an auto router than anything else.
- For the second test I took my time - adding information about trace width, telling him my standards about signal integrity, my routing standards, my stitching standards and so on and so on... And the result was shockingly good. It took 60min to get there but on the first look if it wasn't my design I would have not be able to see what is AI generated and what is coming from me.
Sure this was a simple PCB - but this is V1 Astra - from here it will only get better. I often see threads on reddit where HW guys telling everybody that HW is that complex and that it would take ages that AI would adapt it. But basically this is how SW engineers sounded 2 years ago... And I am pretty sure in 6-12 months AI will also take a big role in my PCB workflow.
Some final thoughts
Looking back at the last years I think the biggest change is not how much better AI got at writing code - I think it is especially how the role of an engineer is changing. You are somehow not anymore the guy making the deep dive, bringing up solutions, banging your head against the wall at 2am..
Now you are more a supervisor of many junior engineers doing the hard work for you and you are only reviewing and thinking about architecture the whole day. For me the challenge is sometimes to still feel like the owner of the code and not just accepting it because "it will be alright".
Some engineers fear that they will be replaced by AI. But I am pretty sure: Good engineers with strong knowledge will always be able to describe a solution in more detail than a non-engineer. And with that AI is not replacing them - it is accelerating them - just adapt this mindset and start building great things faster!