I spent six years building payments infrastructure inside Goldman Sachs. Now I'm rebuilding it from the outside.

SWIFT and ACH were built for humans in banking hours. Stablecoin settlement is instant, but the rails stayed USDC-dominated — so every non-USD corridor secretly routes through the dollar, and global-first companies still lose 3–5% on every currency hop. Meanwhile AI agents have started to book, hire and pay autonomously, and no financial primitive exists for machine-native value flows.

Krypton's bet is that payments, treasury and yield stop being separate systems: one programmable balance that is liquid and earning at once, settling direct pairs with no dollar intermediary, with an agent runtime across all of it. I spent six years inside Goldman building the infrastructure this replaces — the reason to build it now is that the legal rails and the machine demand arrived in the same window.

Delivery solved convenience and took away everything else: no visibility into the kitchen, no personalisation, no control over quality, and a price that stops it ever becoming a daily driver. FoodLabs was built on the insight that this is a compiler problem — any dish, however complex, is a sequence of a few fundamental cooking functions applied in the right order. Once you accept that, you stop building a robot chef.

So we built cheap single-purpose robots that each do one function well, stacked them vertically to use expensive floor space in three dimensions, and put the intelligence in software. And we never raised — we put it into live service instead, which is the part I'd point to. A machine that has to work at service every day is a very different engineering problem from one that has to work in a demo.

Transaction Banking is where the abstractions stop being abstractions — balances moving in real time, against real limits, watched by people whose decisions depend on the number being right. Most of my work was building the instruments the desk steers by, and I kept ending up at the seam between the plumbing and the decision: make the real behaviour of the system legible enough to act on.

Six years of that is also six years of watching exactly where the legacy rails break, and why the breakage is structural rather than incidental. When the regulatory window and the machine-native demand arrived in the same year, I had an unusual amount of conviction about what to build instead. That is Krypton.

Every minute a medical delivery is delayed in traffic, a life is at risk. The bottleneck in emergency care isn't the medicine or the doctor, it's the road between them — and an autonomous UAV doesn't improve that system, it routes around the part of it that fails.

The part I'd point to is the team rather than the airframe: twenty-plus engineers alongside five doctors, because a medical delivery drone designed without clinicians in the room is a toy. Holding that together on student time, with no company and no salaries, is the hardest thing on this page.