Update (June 2026): This is a look back at an early personal experiment. The self-learning engine behind it later became part of the technology I now build at Wakeline, so the free public version described below is no longer online. I have kept the original account intact.
Most of my side projects begin because I want something that does not exist yet. This one began because I was tired of guessing whether the road into Wolfsburg would be clear.
So I built a small traffic predictor for the city and ran it online for free, open to anyone who wanted it. It watched the major roads, learned their daily rhythms, and warned of a jam only when it was genuinely confident. None of that sounds unusual until you look at how it learned.
A System With No Training Phase
The engine behind it was unsupervised. I never handed it a clean, pre-processed dataset and waited for a training run to finish. It learned on its own, directly from the live traffic signal, and kept refining its understanding of each road as it observed it. There was no separation between a training phase and a running phase. It was always doing both at once. This is the property researchers call continual, or online, learning, and it is the thread that runs through almost everything on this blog.
It was also deliberately general-purpose. The same engine once learned to play a video game before I ever pointed it at a road, and I later set it loose on a freeway interchange in Los Angeles. The task changed. The mechanism did not.
Cheap Enough to Give Away
Because it updated a little at a time rather than reprocessing years of history on every cycle, it cost very little to run. It sat on a modest cloud setup and barely registered on the bill. That efficiency was the only reason I could offer it to Wolfsburg for free. A system that demanded a fresh training run on a GPU cluster every week would never have survived as a free public tool.
It Never Stopped Improving
It was never really finished, which was the whole point. I kept adding roads and routes to widen the coverage, and the algorithm itself kept getting refined behind the scenes. The fun part was the daily report card. Every morning the city showed me whether the previous day's change had actually helped.
Why I Care About This Now
The Wolfsburg predictor started as a way to smooth my own commute. It became the clearest early proof of an idea I have since built a company around, and eventually it stopped being a personal toy at all. At Wakeline we build continuous learning systems for environments far harsher than a ring road, starting with day-ahead electricity prices, where the patterns shift constantly and a wrong forecast costs real money rather than a few minutes of your morning. The traffic tool taught me the shape of the problem on a small and friendly scale. The grid is where the same idea has to hold up under pressure.
The story of how it all started, on my own A39 commute, is here in German.
