A Complete System in 60 Hours
In approximately 60 hours, I was able to develop, on my own, a complete ETCS/ERTMS system (RBC, EVC, LEU, JRU, etc.). For those unfamiliar, this is a safety-critical system. For obvious reasons related to company size, timeline, and practicality, the system is not certified SIL4 as it would need to be for real-world deployment.
Is it certifiable on suitable hardware? I believe so, but not under current certification procedures. The goal of this work is therefore, among other things, to move forward in the development of safety-critical systems built with AI by accumulating experience, identifying the weaknesses of this approach, and imagining solutions.
As an aside, shortly before this, I had "only" managed to develop an ERTMS ATO (Automatic Train Operation) in 6 hours.
These are projects that typically take years with large teams, even if we ignore the workload dedicated purely to safety, quality, or management. Assuming we keep only that workload and replace a team with a single engineer, the consequences are staggering. I won't go too deep into my personal opinions here; I'll simply present the facts in the form of questions/answers... and more questions.
1. Is the system close to production, and how do we know?
A legitimate question, considering that even reviewing the produced code (>55,000 lines) would be difficult for a single person in a reasonable time.
It turns out that a paradigm shift is required. I didn't really develop all of this alone. I did it with agents to whom I assigned tasks; those agents created other agents with tasks, all instructed to cover a set of requirements, test them, and prove that coverage — just like a human team would.
Proof of coverage exists and can be verified through complete documentation and detailed traceability in accordance with CENELEC SIL4 requirements.
That said, the project is probably not 100% finished. There are still bugs, but that's simply the normal life of a project — except everything is happening at hypersonic speed.
2. How do we know it's not just "paint"?
Having worked on these kinds of projects all my life — and having had the wool pulled over my eyes by humans — I'll answer simply: hallucinations and "lies" did occur in this project. It's fundamentally an alignment problem, exactly like with humans.
If you ask an engineer to finish a project at all costs, there's a high chance they'll deliver something poor — or even dangerous. Just like with humans, you must constantly realign objectives and perform double or even triple checks.
This can be done either using the same model in a separate session or by using different models. What matters is the role you assign to the agent. If you assign it a role of critic, controller, or verifier, it will find what didn't work, what may be false, or what is inconsistent.
3. Would you ride a train dependent on this development?
As it stands, no. And that's where certification comes in.
Humans are no less prone to failure than the coding agents I used (based on Anthropic's Claude models). Certification processes exist precisely to prevent human error through a systematic and systemic approach to safety issues.
If the same procedures are applied to this project, you get the same results across all critical aspects. Moreover, since it's much easier to create a very comprehensive test suite, I genuinely believe the quality could be higher.
4. What is it like to develop such a complex system with AI?
There is absolutely no procedural difference from what would have been done with a human team.
The difference is that a single person must be able to understand all aspects of the development and have 360-degree technical competence. The typical profile is either a systems engineer who masters development, or a developer who deeply understands the system.
Is a manager needed? I've always considered managers useless — today, I'm certain of it.
5. Are you going to get rich from this?
You'd have to be naive to believe that.
Sure, people might say I can do a couple of things better than others. But I think at least 10% of the engineers I've met in my career could do the same — maybe even 20%.
I don't have a crystal ball, but I'll ask a simple question: if 20% of engineers can produce something that used to require at least fifteen people over one or more years, what will happen?
And even if things go well for that 10–20%, what will happen to the others?
6. What's your business model in making this public?
I don't have one.
Not long ago, someone almost insulted me for not knowing how to "make money." Maybe. (But I'm still here.)
For now, I'm taking stock of things and keeping a bit of an edge.
More than ever, being an engineer has become a race — an increasingly crazy one. The image that comes to mind is an Uber delivery driver who, instead of a scooter, now has a spaceship... but still gets paid per delivery.
7. What do people who already know about this think?
At first, people flattered me: "Very few people could do what you did!"
Then it shifted from "Victor developed it" to "Claude developed it," and from "that's impressive" to "anyone can do it."
People get used to everything quickly.
Individually, some think they're untouchable — on the winning side — while others are anxious about what comes next.
The "winners" have always annoyed me. And honestly, succeeding at something that seemed unimaginable just months ago didn't reassure me.
We should approach what comes next with humility. But do we still have time to reflect?
This post is part of the SS026 project journal — a complete ERTMS/ETCS system built collaboratively between a railway signalling engineer and Claude, Anthropic's AI assistant.