I build for the people with the least room to absorb a mistake — kids, anxious parents, and now patients.

I came to Google with a Master’s in CS and joined, almost by accident, one of the earliest efforts to build for families and kids — well before anything was public. We were a handful of engineers asking what Google should even be for a child. I built experimental family products: location sharing, photo sharing, the small shared infrastructure other teams quietly picked up. I also worked alongside the team building Google’s authorization system — the one later described publicly in the Zanzibar paper — because every question we touched was a question about who is allowed to know what. I learned fast that this domain is unforgiving: privacy, the law, and a child’s development all sit in the same decision.

When Family Link launched, I became tech lead of the team building real-time family location. That sounds tidy; it wasn’t. We were the first team at Google to try to deliver fresh location on a stack designed to be passive and battery-frugal — a benchmark the system was never built to hit. So we measured everything, and I made the unpopular, data-driven UX calls: what “live” honestly meant, what we’d promise a parent, what we wouldn’t. I got promoted to Senior in the middle of it. The lasting lesson wasn’t the latency number. It was that a parent staring at a map at the wrong moment deserves an answer you can stand behind.

Then came the harder problem: extending supervision to accounts that already existed — teenagers with years of their own email, photos, and life. A colleague called it “launching the product again,” and that was about right. The sensitive part was that supervision attached to an existing account can become a surveillance primitive if you’re careless. So I designed a two-party, two-device handshake: both people present, both consenting, on their own devices, and either able to walk away — paired with a transparency app so the supervised person always saw exactly what was visible. That same architecture became how Google met GDPR for children’s accounts.

I started managing here, and grew a team from three junior engineers to eight — every one of them promoted. By then I owned Family Link’s engineering, and the product had a problem it didn’t want to admit: it had gotten big, and reliability hadn’t kept up. Trust was eroding. So I did the thing nobody thanks you for proposing — I pitched executives to halt the feature roadmap for roughly three quarters and rebuild the foundation. There was no playbook, so I wrote the exit criteria myself. Complaints fell sharply even as the userbase grew several-fold. I set engineering practices that outlived my team, across about ninety engineers, and was promoted to Staff. I also owned privacy in one of Google’s most scrutinized spaces — including verifying, not assuming, that child accounts kept no server-side location history.

I moved to Zürich and became a manager of managers across the Assistant-to-Gemini stack. I chartered and rebuilt teams, defined what “production-ready” had to mean for us, and drove a platform consolidation that ended up adopted by hundreds of engineers. My teams shipped the on-device work behind the natural-conversation launches that became Gemini Live. But the part I’m most honest about is the unglamorous half: rebuilding a struggling team, winding down my own team’s scope with dignity when the strategy changed and placing every person well, holding morale through layoffs and the Bard-to-Gemini whiplash. As a senior lead in a ~150-person org I ran recruiting, budget, and calibration. Leadership is mostly what you do when the win isn’t yours to claim.

At YouTube I set the Android platform strategy for apps more than two billion people use. The hard part isn’t the rewrite; it’s that platform decay never wins a quarterly fight on its own. So I made the decay legible — priced it — to win the investment to fix it. I led account-scoping that keeps child, teen, and private-session data strictly isolated, and an AI-assisted approach to migrations otherwise measured in engineer-decades: agents executing a crisply specified contract under human review. The hard part, it turns out, is writing that contract precisely enough that a machine can be trusted with it.

Now I’m pointing the same instinct at the highest-stakes place I can find. I’ve been building toward trust-critical AI hands-on — with humans in the loop exactly where it counts, and accountability that stays with the builder rather than diffusing into a system no one will answer for. The people I’ve always built for are the ones with the least slack, the ones with the most to lose when a system gets it wrong. It’s the conviction I started with, now with the volume turned all the way up. I’m looking for the right place, and the right people, to keep doing the work where it matters most.