How Korso started

A summer on the factory floor#

In the summer of 2025, I took a job at General Motors in Michigan. The assignment was to help deploy Fanuc robots on the assembly line and automate the sequencing of shock absorbers for various vehicles. It was promised to be technical, innovative, and hands-on work. However, halfway through the summer, I realized I was barely touching the robots.

Most of my days were spent on the stuff around them. Figuring out which vendors could actually supply the components we needed, chasing suppliers that were late, and trying to coordinate a mess that was spread out through several teams.

The automation problem on the floor was basically solved (the robots were capable of all that were thrown at them). However, all the coordination holding it together was still running on emails, spreadsheets, and someone's memory. We'd spent years automating the physical work, and yet no one on our team had touched any of the operational layer.

At my end-of-internship presentation, I was confident enough to say that I'd be in San Francisco next summer as part of Y Combinator.

Getting the team together and Figuring out the Problem#

Back at Penn, I pitched the startup to Daichi, my roommate. No idea yet, but a sector and a problem. He'd worked as an engineer at a semiconductor company and got it immediately. Soon, he brought in Alex, a childhood friend of his who possessed the ML knowledge that we lacked.

We talked about it for a while, then decided to stop talking and go find out if we were right.

Before we built anything, we wanted to see the problem for real. So that winter, we just went to factories. Not to pitch nor demo, but to figure out if we were right. We'd show up, ask if we could watch how things worked, and try to understand where the day went sideways. What does a procurement manager actually do all day? What makes an operations lead stay late? What falls through the cracks when things get busy?

By the end of it, we were sure of our idea. Every place we visited had the same problems. The ERP was full of records, yet they came in through the inbox. And somewhere between the two, a person was doing the routing: copying things, chasing replies, escalating what they couldn't resolve, keeping the whole thing from falling apart. By this time, LLMs were getting popular and we realized one thing: basically everything I did with my job at GM could have been automated.

Chronos#

Our first answer was big. We'd build an AI-native ERP from scratch. We wanted to rebuild the whole operational stack with AI at the center. We called it Chronos.

It got us into Y Combinator.

What customers actually told us#

Here's the thing we kept hearing: manufacturers liked what we were working on, but they didn't want to replace their ERP.

That's not a small ask. Replacing an ERP means migrating years of data, retraining everyone, and betting the operation on a new system while it's still finding its footing. Even companies that hated their current setup weren't ready to do that. The ERP, for all its problems, was load-bearing.

What they actually wanted was help with everything happening around it. The coordination, the exceptions, the follow-up. If we could connect to their existing systems and take that work off someone's plate, it would be something people wanted.

What we built#

We rebuilt Chronos into something different. Instead of replacing what manufacturers already had, we built a layer that works on top of it. That's Atlas^↗.

Atlas connects to the ERP, the inbox, whatever other systems the operation runs on. It watches what's happening, figures out what needs to happen next, and either does it or brings a human in at the right moment. The coordination work that used to live in someone's head and inbox now runs 24/7 and remembers everything it's ever done.

The full thinking behind why we built it this way is here: Korso thesis^↗.

Automation is easy for those at the frontier, but we're trying to fill the gap for manufacturers that can't afford to build their operations from scratch. The answer to this evolves every day, but for now it's Atlas.