I raised with no idea, deck, or company -- and earned our way into the problem.
August 30, 2026
I've always been more of a practical person.
I like experimenting, building things, seeing what breaks, and changing my mind based on what I learn. Especially with a startup, I'd rather discover the problem by doing than spend years convincing myself I've found the perfect idea.
When I raised Cayu's pre-seed, I basically had no company.
No product. No real idea yet. I actually had to incorporate the company, open a bank account, and get everything set up after signing the fundraising documents.
What I did know was where I wanted to spend my time: bringing AI into the physical world -- manufacturing, supply chain, energy, logistics.
That's where I had already spent most of my career: at McKinsey, at C3 AI, and during my PhD through internships with Western Digital and Boehringer Ingelheim's manufacturing team.
So I started experimenting in January 2025.
My first beachhead was trucking. I happened to meet one of my first clients at Manifest 2025, and I had also worked with some of the largest transportation companies during my time at McKinsey, so I understood the problems reasonably well.
Technology penetration in the industry had historically been low, and I thought LLMs might finally change that.
I built copilots for logistics companies and got them into production.
I learned a lot.
I also learned that the market was smaller and more fragmented than I expected. Relationships and insiders mattered enormously. Better technology alone wasn't suddenly going to change how the industry bought software.
At the same time, tools like Claude Code started commoditizing many of the easier use cases -- recruiting, TMS workflows, and others. I started seeing many, many copycat startups.
And in many cases, buyers simply couldn't distinguish between an S-tier product and a C-tier product. The quality of the underlying technology was not enough to create differentiation.
But trucking gave me something much more valuable: my first real experience operating LLM-based systems in production.
This period also coincided with several experimental CTO setups not working out.
Many people wanted to join once product-market fit existed. Fewer wanted to go through the messy learning process required to find it.
Then I started talking more seriously with a longtime friend. We had known each other since our International Olympiad days -- I competed in IChO, and he won a Bronze Medal at the International Mathematical Olympiad.
He had already had an exit from his first startup in California and was working at a big tech company, but he was eager to build another company.
We started discussing how LLMs were changing the creation of digital assets and decided to experiment together while I still had runway.
We went after what I considered one of the hardest digital assets to create: software.
Text had become easy to generate.
Images were becoming easy.
Production software was different.
So we spent months building coding agents and putting them into production.
We launched a "vibe-coding" platform to build enterprise-grade software.
In ~60 days, we reached over $200K ARR. Real ARR -- not a $99 discount counted as $100 of annual revenue.
But we also noticed that building software was going to become a thing of the past. Agents were the future.
That changed my view of agents completely.
The impressive part wasn't generating something in 30 seconds.
It was what happened over the next 30 minutes, 3 hours, or 3 days.
Agents had to maintain context, use tools, recover from failures, retry, ask humans for help, remember what had already happened, and keep working toward an outcome.
Eventually, something clicked for us: This wasn't a coding-agent problem.
Long-horizon agents were going to become important far beyond software -- across finance, manufacturing, supply chain, energy, and eventually personal work as well.
When we started taking those learnings back into industrial enterprises, we kept encountering the same underlying problems: state, recovery, approvals, policies, budgets, auditability, memory, and continuity.
At first, we solved them separately inside each application.
Eventually, we realized we were rebuilding the same infrastructure over and over again.
That became Cayu Runtime.
So yes, over ~20 months, Cayu has pivoted several times:
Logistics copilots → production coding agents → infrastructure for long-running enterprise agents.
From the outside, those can look like three different ideas.
From where I sit, it feels much more like a series of experiments that kept taking us one layer deeper.
I'm actually glad we didn't start with an "agent infrastructure" thesis and then go searching for reasons it should exist.
We earned our way into it by building things in production.
And for me, that's one of the most enjoyable parts of building a startup: you get to keep experimenting until you find a problem meaningful enough to spend years of your life on.
P.S. Right when we were testing the Cayu Runtime thesis, a longtime friend from Bain and HBS -- and a former CFO of a ~$2B revenue energy company -- reached out to me.
He ended up joining us too.
That's a post for another day.