Improving the quality of personalization on Google AI Mode. Through careful feature engineering and precise experimentation, I shipped improvements to the LLM's responses and its handling of saving and applying user data.
The problem space
Personalization on Google AI Mode is a careful balance of latency and quality. Users expect fast, accurate responses that adapt to user info. Improvements require measuring statistically significant metrics movement during prototyping and experimentation, and after launch. The problem space involves identifying gaps, developing end-to-end features and improvements to address core user needs, landing features after extensive quality analysis, and justifying the entire process with evidence.
What I owned
- Measured the impact of shopping vertical AI features on key Search metrics, which drove high-level decision making.
- Hillclimbed quality on personal context for shopping queries, by improving tool calls and prompt logic.
- Improved memory triggering on multi-turn queries, by utilizing a novel method to provide previous-turn context while conserving token usage of the lightweight LLM classifier.
- Identified and fixed a key feature gap with gift queries, by improving response quality and personalization accuracy.
The hard part
[The one problem that took the most thought. What made it hard, what you tried first, and what finally worked.]
The result
[What changed because of the work. Use numbers only where they're real and cleared to share.]
- [What this number measures]
- [#]
- [What this number measures]
- [#]
- [What this number measures]
- [#]
What I'd change
[With hindsight, one or two things you'd do differently, and what you took from them into your next project.]