A2UI Design Principles
Moving from responsive design to "Agent-to-User Interfaces" (A2UI). Why generative UIs must prioritize observability over aesthetic minimalism. — Aadhar
Back to work
Making Gallery search work for people
Gallery search worked when you knew the filename. It failed when you only knew what the photo was of.
- 10% → 18%
- search adoption, an 80% lift
- +40%
- average query length
- +14%
- Gallery click-through
How people searched
People do not search their own photos the way they search the web. They search for an occasion, a person, or a place: a sentence rather than a keyword. Adoption sat at 10% because the product only rewarded queries the system already understood.
Query decomposition
We used an LLM for query decomposition, so a natural-language sentence could be broken into the attributes the index held. Average query length rose 40% once it worked. People started asking for what they meant instead of guessing at keywords.
Putting the right result first
Retrieval alone did not move click-through. AI-powered top hits and most-relevant ranking lifted Gallery click-through 14%, because the right photo needs to be in the first row, not merely in the result set.
Gallery and Health
The pattern carried into Samsung Health. GenAI-led health scores and wellness tips lifted app usage 10%, and search and discovery across Gallery and Health together delivered an 8% lift in weekly active users.
Working at scale
This shipped to a device base measured in hundreds of millions. Adoption moves slowly at that scale. Lab accuracy does not tell you whether people changed how they search.
Related writing
The Bottleneck Was Never the Model — And It Just Moved
The companies pulling ahead are not waiting on a smarter model. They noticed the bottleneck was never the model — and that since they fixed it, it has moved twice more. — Aadhar