Life
Coupang and the Enterprise AI Turn
When I left Google in 2024, I joined Coupang. On paper it was a year and change. In practice it was the most compressed stretch of my career — I started four or five distinct things, made them real, and handed most of them off. Looking back, that was the actual job, even though nobody wrote it down that way.
Traffic
I started in infrastructure, on a problem that sounds narrow and isn't: what should a system do when it is overloaded? The naive answer is "fall over," and at e-commerce scale the naive answer is expensive. I worked on adaptive load shedding — dropping the right requests instead of all of them — and the moment you take that seriously, it stops being a feature and becomes a worldview. Load shedding grew into rate limiting. Rate limiting grew into a broader traffic control program: who gets capacity, under what policy, decided where in the stack.
Then I gave it away. Once the program had a shape, a roadmap, and momentum, I handed it to another L7 IC and moved on. At earlier points in my career that would have felt like losing something. By this point I had made peace with a pattern I seem to repeat: I am most useful in the zero-to-one stretch where a problem doesn't have an owner, a name, or a budget yet — and the right move at the end of that stretch is to put it in the hands of someone who will run it for years.
Globalization, again
Next came globalization — g11n platform specs for the company. This was a strange homecoming. I had spent 2016 to 2018 building Rosetta at Uber, and here I was nearly a decade later writing down how a different global company should structure translations, locales, and the platform underneath them. The difference was that this time I had already made the mistakes once. Specs are an underrated artifact: the code gets rewritten, but a good spec shapes how a company thinks about a problem long after you leave.
The agentic turn
Then the center of gravity of my work shifted, the way it was shifting everywhere in 2024: toward AI. I ended up leading the enterprise AI effort — equal parts infrastructure, workflow design, and culture.
The concrete pieces, roughly in order: we got Codeium (later Windsurf) into engineers' hands, which sounds like procurement but is really a security, legal, and adoption project wearing a procurement costume. We built a centralized MCP gateway so AI tools could reach internal systems through one governed door instead of a hundred ungoverned ones, with audit logs on everything — if you cannot answer "what did the AI touch and when," you do not have an AI platform, you have a liability. And we built Emux, our agentic execution platform: the substrate for actually running agents against real company systems rather than letting them loose ad hoc.
The project that taught me the most was using agents to rewrite employee data to take risk out of the company — de-identification is the term, I think; the compliance people had a more precise one that never stuck in my head. The interesting part wasn't the redaction itself. It was the inversion: we were using agents to make the company safer from data risk, at the same time everyone was asking whether agents themselves were a data risk. Both things were true. The answer to both was the same machinery — gateways, permissions, audit trails, evidence.
What the year was actually about
I wrote a whole essay series out of this period, but the honest summary is shorter: a big company cannot adopt agentic AI by buying licenses. Somebody has to build the rails — the access layer, the audit trail, the execution substrate, the observability — before agents can be allowed to act on anything that matters. That somebody usually has no job description, because the job didn't exist the year before.
Doing that work is where my deployment-control thesis stopped being an opinion and became lived experience. It is also what made the next move obvious: if the rails were the interesting part, the most interesting place to build them is where the AI's decisions carry the most consequence. That meant moderation. That meant Roblox.