Make models fast and cheap by picking speculative decoding, prefix caching, sharding, and quantization bets, proving them with load tests and live experiments, and landing them. Get to know every new Gemini revision better than its release notes do, analyze what changed in its architecture, behaviour, and cost to turn that into a serving topology that gets the most quality out of every TPU chip. Build the guardrails that let us move fast without breaking Search: configure checks, pre-submits, monitor and fast rollback, fix the root cause, not just the alert.
Qualify new checkpoints, shape evaluation and autorater strategy, and explain why the new revision is suddenly slower, and what steps needs to be taken to improve it. Write the design documentation people read, and raise the bar in design and code reviews.
Job description provided by Google.
View the original posting