Designing the AI That Feels Alive
Turned Qwen's static docs into interactive showrooms — first proof moment dropped from 60+ minutes to under 2 minutes, with post-launch traffic at ~2× the pre-launch baseline.
Design principleAI systems need visible cognition, not just outputs — I design to make model state inspectable.
- Role
- Sole UX designer — research to production code
- Timeline
- 4 weeks · July–August 2025
- Team
- Me · 2 supervisors · 2 PM · 1 engineer
- Owned
- UX strategy · 4 showrooms (UX + visual) · design system templates
The Problem
The first hour was killing trial conversion
The docs explained everything. But feeling the model meant configuring, running samples, and interpreting output alone — a loop that routinely stretched past an hour. Most trial users left before reaching the moment of value. Enterprise decks faced the same wall: descriptive, not convincing. So I shifted the product from documentation to proof.
Before
Static documentation and generic chat — previous experience (screen recording)
How might we
Make model capabilities visible, testable, and trustworthy — within minutes?
Decision 01
I replaced documentation with market-specific showrooms.
Instead of improving the documentation, I designed 4 market-specific showrooms that let users experience a working version of their own future product. Companionship, psychotherapy, character cloning, IP licensing.
Before — Generic chat
After — Therapy room
6 apps, 40 + comments — every competitor felt like another ChatGPT. The answer was market-specific showrooms: one vertical per room, built for the evaluator who already works there.
Users don't believe descriptions. So the first message had to prove the capability.
What makes AI experiences hard
My role was to turn invisible model behavior into visible product surfaces.
Users don't only judge the output. They judge whether they understand what the system knows, why it responds, and how much control they still have.
So I designed 3 forms of visibility:
Memory visibility
The system recalls and updates personal context in the flow.
Analysis visibility
The system shows what it understands while the conversation continues.
Implementation visibility
The system exposes prompts, YAML, and constraints beside the live demo.
Decision 02
I designed each room to prove one capability in 60 seconds.
Three model strengths crammed into one chat window. None of them landed. So I split them across rooms — one proof moment per room, legible in 60 seconds, no explainer text.
Showroom → One proof → In-product behavior
Romance Room
Long-term memory
In experience
Character recalls conversation specifics across sessions
Astrology Room
Real-time memory updates
In experience
Live constellation profile updates mid-conversation
Therapy Room
Real-time analysis
In experience
Expert panel surfaces conversation themes as you chat
Memory in the UI
Most AI products treat memory as invisible state. I made it a first-class object in the UI — because in B2B evaluation, what you can see is what you can trust.
Each room required its own interaction language. Below, a deep-dive on the Romance Room — designed to sustain engagement through long-term memory and parallel narrative threads.
Astrology Room
Real-time memory updates
A personal constellation file updates during conversation — memory becomes transparent and inspectable.
Therapy Room
Real-time analysis
A live panel surfaces conversation themes — users see what the system understood, not just what it said.
Decision 03
I made demos emotional for users and inspectable for builders.
Inspiration and Continue Response guide users to the wow moment. A slide-out drawer keeps YAML and prompts next to the live demo.
The question shifts from “can your model do this” to “how fast can we ship.”
Inspiration Response
Three reply options — action, emotion, expression. Guides without breaking flow. Feels like gameplay, not messaging.
Continue Response
One tap extends the story from context — long-context reasoning, no effort required.
Code drawer, not console.
YAML specs, prompts, and constraints slide open beside the live demo — no context switch. Evaluators can inspect implementation without switching context, and clone the template as a reusable starting point for their own product.
How I Worked
AI changed how I shipped the work, not just how I made assets.
AI compressed the distance between strategy, visual direction, motion, and implementation. I used it to stress-test decisions, generate visual directions, prototype motion, and deliver production-adjacent interfaces that engineers could merge with minimal revision.
Inspired by Love and Deepspace, I used Wan, Kling, Dreamnia, and SeeDance for visual identity. Interactions built with Cursor and Claude Code — all in four weeks.
Product Showcase
Live interactive prototypes
Open full page for best fidelity.
Romance showroom
Long-term memory and emotional pacing in one flow.
Additional Contribution
Full refresh of the Qwen Character SaaS console.
I delivered an end-to-end update to the Qwen Character admin — spanning API surfaces, Studio (Applications, Workflows, Knowledge Base, Characters), and the nested flows underneath. That included secondary screens teams rely on in production: empty and error states, and analytics views for operational data such as invocation metrics and call volume.
Design framework
Adoption
Spark Design templates — adopted B2B design system page
If the frame is empty, the host blocks embedding — use Open in browser.
Impact
Shipped. Converted. Adopted.
3
Showrooms shipped
romance · astrology · therapy · character
~2×
Model traffic
token & call volume vs 4-wk pre-launch avg
49+ min → <2 min
Onboarding
static docs to first proof moment
3 weeks
Research to production
intern project, shipped
Takeaway
What I learned
AI products do not sell themselves through capability lists. They need proof moments, visible cognition, and inspectable systems. The designer's job is to translate model behavior into experiences people can feel, trust, and build from.
Memory transparency
Constellation file makes memory readable · not a silent black box
Analysis visibility
Therapy rail shows what the model understood · not just what it said
Developer inspectability
YAML + prompt exposed in code drawer · inspect before you build
Emotional boundary
Therapy room = analysis demo · no clinical claims implied
Design principles
Design is the translation layer.
In AI products, the hardest problem isn't the model — it's helping people imagine what to build.
The best demo is future-self proof.
Show a working version of their product, then let them clone it.