Leading Google's Assistant into the Gemini era

I came to Zürich in 2022 to lead assistant engineering, and the remit grew fast. Generative AI arrived mid-flight, Assistant became Gemini, and the work I owned landed at the center of the company's top priority. What I led wasn't a single team or a single layer — it was a portfolio that ran the length of the stack:

On-device intelligence

The assistant running locally on phones, watches, cars, and the home — fast, private, and useful even offline.

Server fulfillment

The layer that turns an understood request into a real action — the difference between answering and doing.

Gemini Live

Natural, free-flowing conversation you can interrupt and redirect, instead of one rigid command at a time.

Device ↔ cloud handoff

One conversational protocol so a session moves between phone and server seamlessly — light work local, heavy work in the cloud.

Reliability

The readiness and quality of assistive experiences end to end, across surfaces and conditions.

Stack unification

Collapsing parallel Assistant and Gemini stacks into one, so a capability built once works everywhere.

The through-line was unification. Years of Assistant and the new world of Gemini had grown into parallel stacks; my mandate was to make them one — so a capability built once worked everywhere, and a conversation could begin on a watch, continue on a phone, and reach into the cloud for a heavy action without the user ever feeling the boundary.

The best handoff is the one nobody notices: light work on the device, heavy work in the cloud, one conversation throughout.

The structural bet underneath it was consolidating custom, siloed systems into one core-supported platform — adopted by roughly 800 engineers across Google, cutting redundant infrastructure cost by as much as half, standardized on Kotlin so the same code could serve both server and device. And the on-device half kept its own discipline: every capability fit inside per-device-class budgets for latency, battery, and memory, because running models where your data lives is the strongest privacy promise a consumer AI platform can make.

As capabilities multiplied across surfaces, quality threatened to become anecdote. I chartered a reliability and quality effort that turned “is it ready?” from a judgment call into a measured bar — so an assistive experience shipped when the numbers said it was trustworthy, not when the demo went well.

Correctness

Does it do what was asked? Measured per capability, with budgets attached.

Reliability

Does it do so every time — across devices, locales, and network conditions?

Latency & health

Inside the device-class budget, without degrading the system around it?

Debuggability

When it fails, can an engineer see why? Failures you can't explain are failures you can't fix.

The pillars fed a continuous-readiness bar with launch gating. On Family Link I had learned this lesson after a crisis; here we built it in before one.

I led this portfolio by managing managers and senior staff, not by touching every change — and I was one of the pillars of the ~150-person organization around it, helping run its operating rhythm and its performance calibration. That breadth mattered most when the ground moved: this era brought layoffs, the Assistant-to-Gemini pivot, and reorganizations.

I did the unglamorous parts well: rebuilding a struggling team's leadership end to end, divesting systems we shouldn't own, winding down scope my own teams had built when the strategy genuinely changed, and protecting morale and retention through the uncertainty. What I learned leading at this scale: set the direction in writing, measure what “good” means, and when the plan changes, say so plainly and early. Teams forgive a changed strategy; they don't forgive a hidden one.