Project spotlight

Sage+: Generative AI
for Knowledge Discovery

Amazon — 2023–2025 — Senior UX Designer

AI / Generative AI 0→1 Product Design Leadership Mixed-Method Research End-to-end UX 3 products + Slack + AtoZ

In 2023, knowledge discovery at Amazon was a known pain point for software engineers. Finding a reliable answer meant knowing where to look, who to ask, and whether the information was still current. The Knowledge Tech team set out to build their own LLM-powered service to fix that. I was brought on as the sole UX designer to build the experience around it, from the ground up, across five surfaces over two years.

Design impact

+47%
Unique active Wiki users (Mar→Oct 2025)
+39%
Answers served in the same period
64%+
CSAT maintained at all times
+10%
Positive feedback lift from a single design change

My approach

01 — Establish
Foundational research
No internal AI UX patterns existed yet. I built a trust and expectations framework from scratch that guided every integration that followed.
02 — Build
Design in lockstep
The underlying model was evolving in real time. I adapted the process so design and validation moved alongside development, not after it.
03 — Validate
Continuous measurement
Established a quarterly CSAT mechanism from day one. Gave leadership a consistent baseline and surfaced improvement opportunities with each launch.

What I designed

  • Sage (Q&A) The first integration. AI responses surfaced within two minutes of posting a question. Established the trust research and visual identity that carried through every surface that followed.
  • Internal Search Implicit and explicit AI alongside search results. Designed a non-disruptive loading pattern and a user-controlled toggle after engagement data revealed a clear split in user preference.
  • Slack Explicit chat experience in a conversational context. Defined emoji language patterns, scoping controls, and response framing tuned to the way engineers use Slack for in-the-moment problems.
  • Wiki All three modes: implicit, explicit, and chat. Made the case to delay launch until page-level scoping was ready, a decision that significantly improved the experience for complex engineering documentation.

Selected moments

Research → design → metric
A third option that moved an annual goal
Positive feedback had plateaued. Qualitative research revealed that binary good/bad feedback was too coarse for AI responses that were partially correct. I designed a "somewhat helpful" option, which produced a 10% lift in positive feedback and helped the team meet their annual target.
Range — beyond the screen
DevCon: guerrilla research at scale
When the team was invited to Amazon's internal developer conference with 7,000+ attendees, I produced the booth materials and turned it into a light research opportunity. Informal survey posters sparked conversations and generated qualitative signal the team couldn't have collected any other way.