My first real job in Silicon Valley was inside a lending company in San Jose. I was hired somewhere between marketing and engineering, which suited me, because I was somewhere between marketing and engineering. I automated their loan decisions, built a portal so customers could stop calling us, and watched something click: when you write software against a business metric, the metric moves. Code was never the product. The number was the product.

So I did what everyone in the Bay Area does with a good lesson. I over-applied it. In 2016 I co-founded Carnectify, a two-sided marketplace connecting banks and dealerships for vehicle financing. The software worked. Real-time underwriting, clean integrations, everything automated. The market shrugged. Two-sided marketplaces need both sides to show up at once, and neither side shows up first. I got to talk it through with Andrew Chen, who was running rider growth at Uber then and knew more about marketplace cold starts than anyone alive. He was generous with his time and honest about the odds, and told me that if I pulled it off, he'd want to write the check himself. He never had to. We couldn't crack it.

What I took from that failure shaped everything after: distribution is not a feature you add. It has to be engineered into the business itself, as loops, not launches.

I wanted to test that idea somewhere nobody would look for it. A friend's used-car lot, a few dozen cars, the epitome of an inefficient business. Perfect laboratory. Over the next seven years I ran hundreds of experiments there. Most failed, quietly and cheaply, which was the point of building the infrastructure first. The ones that worked, compounded. QR codes replaced window stickers, so anonymous lot visitors became retargeting audiences. Every sale was fed back into Meta's ad engine, so the algorithm learned who buys cars instead of who clicks. Financing moved in-house, so every customer became the next customer. Four years in, that little lot was a top-100 independent dealership in California. Seven years in, it had passed a hundred million dollars in cumulative sales.

The marketplace failure taught me the theory. The car lot proved it in the least forgiving market I could find.

Since then the contexts have kept getting stranger: aerospace, edge AI, dual-use technology, rooms with presidents and sovereign funds. The job has stayed exactly the same. Build something that works, then engineer the system that gets it into the world's hands, and be willing to be wrong most of the time, cheaply, until you're right in a way that compounds.

That's the story. The photos on the front page are just what the job looks like some days. The other days, it looks like an IDE.