Twenty-five attempts.
Two down so far.

Back in November of 2025, I attended a business forum in Miami with one of the sickest lineups I've ever seen — Serena Williams, Jamie Dimon, Ken Griffin, Lionel Messi, Jeff Bezos, to name a few. One thing Jeff Bezos said really stuck with me: he was talking about real estate development, and how it shouldn't take so long for developers to figure out whether a property can actually be built on. The data already exists. You should be able to get an answer in seconds.

That got me thinking — if the data already exists, what else can you do with it? So I started looking at New York City specifically. At first I was underwhelmed: the government websites were slow, inconsistent, scattered. But then something clicked — instead of seeing terrible government websites, I started seeing a technology problem. That eventually became Underlink.

What I built

Underlink is an AI-powered lending intelligence platform built for NYC private lenders. Instead of jumping between PLUTO, ACRIS, DOB, HPD, DOF, ECB, BIS, permits, complaints, and mortgage history, a lender could search one property and have everything surfaced in one place. I wasn't trying to replace underwriters — I wanted to help them get to a quality first impression in minutes instead of hours. Underlink was built to answer one question: "Is this property worth spending more time on?"

What I learned

I thought the hardest part would be building the software. It wasn't — building the product was the easy part. Trying to make a company out of it was the hard part.

The first person I showed it to, someone working in commercial real estate lending, told me:

That one comment changed my positioning — I'd been thinking of Underlink as a full underwriting platform, but it made more sense as something lenders used in the first few minutes of evaluating a deal. I changed the marketing to match. Then I showed it to someone who owned a commercial lending firm. We talked for an hour. At the end he said:

That was hard to hear, but it was one of the most valuable conversations I had. He wasn't criticizing the engineering — he was telling me I hadn't created enough value to justify paying for it. Building something impressive isn't enough. You have to build something people are willing to pay for.

I also started too small. If there are only 100 potential customers and you're trying to land 20 of them, that's an enormous share of the entire market — small markets make everything harder. And as an engineer, being able to build my own product is a huge advantage, but building the software isn't the company. Sales, marketing, finance, customer discovery, distribution, positioning — those are all skills too, and sales especially felt uncomfortable. You can build an amazing product, and then you actually have to go show it to people. That's the whole game.

This project also taught me something about myself: I love messy data. I love taking information spread across dozens of places, figuring out how it connects, and structuring it into something useful. That's probably the biggest thing I'll carry into whatever I build next.

Looking forward

This is attempt #1. I'm completely okay if it takes 25 attempts — not because I expect all 25 to work, but because I expect every one to teach me something the previous one couldn't. I don't know which attempt will be the one. Maybe it's #2. Maybe it's #8. Maybe it's #25.

Underlink didn't become the company I hoped it would. It became something just as valuable: the foundation for how I'll build every company after it. On to attempt #2.

One of the most valuable conversations I had while building Underlink came from a lender who told me he wouldn't pay for it. But before we ended the call, he said something else:

That stuck with me. One of the biggest lessons I'd already learned from Underlink was that market size matters — commercial real estate lending was a relatively small market, and short-term rental investing was significantly larger. That alone made me curious. So I decided to learn everything I could.

What I built

The first thing I did wasn't write code. I asked an experienced investor to walk me through exactly how he evaluated a property, then went and did it myself — manually underwriting around 40 properties over the next few days, watching hours of YouTube, and reading everything I could find. I wasn't trying to build software yet. I was trying to think like the person the software was for. Only after I understood the workflow did I ask myself one question: which parts of this can software actually automate?

That became the project. I built a lead qualification and deal-screening tool for short-term rental investors — an investor enters a property address, and within seconds the platform tells them whether it clears a target yield using property data, revenue estimates, comparable listings, and a full expense model. Instead of pulling information across a dozen tabs and spreadsheets, an investor gets an instant go/no-go before deciding whether a property is even worth spending more time on. I integrated several APIs to make it work, including RealtyAPI, Google Maps, and AirROI. The hard part was never the interface — it was recreating an experienced investor's decision-making process in software.

What I learned

This project reinforced something I first discovered while building Underlink: I can walk into an industry I know nothing about, learn how experts make decisions, break that down into first principles, and figure out what software can automate. Being an outsider actually helped — I wasn't thinking about how things had always been done, I was just trying to understand the workflow well enough to build something useful.

I also carried over a lesson from my time at Hopscotch: your product isn't just your code, it's also the companies your product depends on. I relied on AirROI for revenue estimates. They seemed legitimate, but they were small, and when I emailed them with a technical question, I never heard back. Weeks went by. Nothing. That changed how I think about vendor risk — if your product depends on someone else's API, your business depends on their reliability too.

Then there was the business model question. An investor mentioned that experienced buyers often skip a traditional agent entirely and handle transactions themselves. That got me thinking — what if a platform hired in-house agents, paid them salaries, and split the buyer's commission with customers? Before I got too excited, I went looking for who'd already tried it, and found Prevu. They'd raised around $8 million, operated for about ten years, and were eventually acquired for around $10 million. That immediately caught my attention — if this business model was so compelling, why wasn't the outcome bigger? I started digging into the economics. That was the real lesson: building a product people use doesn't automatically mean you've built a great business. The business model has to work too.

The last lesson was the hardest one. After I showed the MVP to that same investor, he gave me several ideas for how to differentiate it. Technically, I could have built every one of them. That wasn't the problem. The problem was that I didn't actually want to spend the next five years building software for short-term rental investing. I'd enjoyed learning the industry. I'd enjoyed building the product. But I didn't care enough about the problem itself to keep showing up for it. Curiosity is enough to build an MVP. It isn't enough to build a company.

Looking forward

This project reinforced a lot of what Underlink had already taught me — talk to users earlier, understand the market, understand the business model, validate demand before writing too much code. But it also taught me something new: looking back, I don't think this project was ever really about short-term rentals. It was another chance to walk into an industry I knew nothing about, learn how experts make decisions, and figure out what software could take off their plate. I've now done that twice. At some point, it stops feeling like a coincidence and starts feeling like a pattern. I'm less interested in a specific industry and more interested in a specific kind of problem: messy data, fragmented information, complex workflows, figuring out what software can automate. That's the thread I'm going to keep pulling on. On to attempt #3.