The hardest work left in the factory is still done by hand

Software has learned to write, to code and to reason. It still cannot plug in a connector.

Walk into any car plant. Robots weld the body in seconds, then it rolls into final assembly and the robots stop. From there, people take over. They route wire harnesses, push in connectors they cannot see, clip, fasten and check, every sixty seconds, for an entire shift. The same is true for almost everything with a motor, a circuit board or a cable inside it.

Hands make mistakes, so factories pay other hands to find them. In the plants we have spoken with, between one in ten and one in five workers do nothing but inspect and repair other people's work. Even then, decades of research show that people inspecting by eye miss a fifth to a third of all defects. What gets through ends up in a bill that keeps growing: in 2025, the world's thirteen largest carmakers paid more than 57 billion dollars in warranty and recall costs, 63 percent more than a decade earlier.

Automating this work is the obvious answer, and the industry has tried. Tesla went furthest, and in 2018 Elon Musk wrote: “Excessive automation at Tesla was a mistake… Humans are underrated.” Since then, billions have gone into humanoid robots. The ones working in car plants today carry boxes and load fixtures. None of them routes a harness.

The bottleneck is not just the robot. It is what the robot knows. Language models were trained on almost everything humans have ever written down. Nobody ever wrote down how a harness is routed by feel, or what a connector feels like in the moment it clicks into place. That knowledge lives in the hands of millions of people on assembly lines, and it has never been recorded. You cannot scrape it from the internet. You have to go and get it.

We believe the machines that take over this work will learn it from the people who do it today. Getting there takes three steps. First, understand what actually happens at a workstation, step by step, every second of every shift. Second, use that understanding to make the work better, catching mistakes before they leave the station. Third, once the understanding is deep enough, hand the work to machines.

Trenon is built to take all three. We start with the first two, in real plants, today, and we will take the third as soon as the machines are ready for it. A camera, worn by the operator or mounted at the station, follows the assembly as it happens, knows which step is running, and speaks up the moment something goes wrong, at the station instead of at the end of the line. Worn, it goes where the operator goes, including inside the car, where no fixed camera can see; mounted, it watches a bench where the part stays in view. We watch the work, not the worker: we do not identify operators, and we do not score them. A factory only lets you onto the line if you are useful from the first day. That is how we earn our place, and every shift teaches us more about how skilled people do physical work.

Our mission is to take human work out of the factory.

Not the people. The work nobody should have to do: eight hours of staring, the connector pushed in blind, the inspection that exists only because people make mistakes and other people have to find them. America alone may leave 1.9 million manufacturing jobs unfilled by 2033. This work will be automated either way. The question is who builds it, and whether they build it with the people on the line or around them. We are building it with them.

Here is why this is hard, and why we think it is one of the most interesting problems in physical AI. Video from a worn camera never holds still. The operator's own hands hide the very thing you need to see. Two steps can look identical, and the difference between a connector that is seated and one that is not is a few millimeters. The bar is brutal: a system that is right on 99 percent of actions still raises close to a hundred false alarms in a single shift, and operators switch off anything that cries wolf. Nobody has published a system that clears that bar on a real line. We intend to be first, and we are moving fast: we built our first prototype in days, and it is now going onto production lines at US automotive suppliers.

Florian grew up hearing his father, a quality inspector for almost thirty years, talk about the mistakes made on the line. He studied mechanical engineering at TU Munich, built industrial software and AI, then co-founded a company, took it to six figures in revenue within months, and shut it down to do this. Nadir started his first businesses as a teenager, graduated first among more than 2,700 students in his business degree, worked at one of Europe's fastest-growing startups and in venture capital, and went straight to McKinsey, where he advised automotive suppliers, before leaving to build. In under three months we have spoken with more than a hundred people who run, inspect and fix assembly lines in Germany and the US. We are backed by Entrepreneurs First.

The first engineers who join us will decide what this company becomes: the architecture, how we measure, and the standard everyone after them has to meet. If you have made a vision system work outside the lab and still remember exactly what broke, if you would rather spend your first weeks on a factory floor than in a ticket queue, and if you want to work on a problem most people think is still years away, we should talk.

Every robot that ever learns to build a car will have to learn it from someone. We intend to be how it learns.