Align with how people actually think — not how tools are categorized.
Users think in goals, tasks, and situations. The interface starts with natural language input and structures it into meaningful intent dimensions.
Case Study · 2024
Designing for how people think, not how tools are categorized.
Nexa — a search-first web app that helps users clarify intent, reduce cognitive load, and confidently navigate AI tool discovery.
Live · 5M+ monthly visits
Scope of work
- Product
- Role
- Timeline
- Methods
- Deliverables
When I entered the team, the product had just released its first MVP. My role began with understanding its real-world performance — reviewing analytic data, observing user behaviors, and interviewing early adopters.
Lori Cai · Project Design Intern · Jan – Apr 2024
I · Research Process · 1.1 MVP Critique
Although the product successfully aggregated thousands of AI tools, users didn't know how to navigate or express what they needed.
My first task was examining how users actually interacted with the MVP across web and mobile. The critique surfaced five categorical issues — none about the catalog itself, all about the gap between user intent and system behavior.
-
01
Noise
Navigation Overload
The top navigation exposed too much content — creating noise and making it unclear where users should begin.
-
02
Low usage
Ineffective Search
The search bar had low usage. Users didn't know what to type, and the system didn't support vague queries.
-
03
Data ≠ user
Data First, Not User
Filters were generated by popularity — resulting in random, unhelpful options that didn't match how users thought about tasks.
-
04
Click-heavy
Space Constraints
Limited space for showing info, but balancing different user needs was hard. Users had to click into every tool to evaluate it.
-
05
Off-system
Inconsistent Mobile
On mobile, filters and UI patterns were misaligned with the rest of the system — further hurting clarity and usability.
Reframed as an HMW — "How might we create an intuitive, intent-aware discovery experience that helps users start easily, refine confidently, and evaluate tools with minimal effort?"
1.2 · Understand the Industry
2024 — when people stopped searching and started asking.
2024 marked a turning point in how people interacted with AI. As large language models became mainstream, users grew more comfortable expressing needs in natural language — but far less comfortable navigating categories, keywords, or technical terminology.
Most users knew what they wanted to achieve, but not what kind of AI tool, GPT, or API could help them do it. Meanwhile, the ecosystem expanded faster than user literacy: thousands of tools, overlapping features, inconsistent naming, and unclear differentiation.
1.3 · Six AI Product Design Principles
By aligning model strengths with market needs, I shaped the value proposition that became the north star of every design decision.
Every step — intent extraction, categorization, reasoning — should be legible.
The system's interpretation of text is exposed so users can validate and correct it. No black box.
Guide momentum forward instead of presenting dozens of filters.
Light-touch refinement steps, recommended categories, and curated results — designed for forward motion, not analysis paralysis.
User input is often vague, incomplete, or incorrect.
Handle ambiguity gracefully — ask clarifying questions, offer suggestions, never punish unclear input.
Show why a tool matches, what makes it different, how it solves the goal.
Results aren't just listed — they're framed by the user's intent, so each entry feels like a recommendation, not a search hit.
Discovery feels consistent on desktop or mobile.
Layouts and refinement steps follow shared logic even as presentation adapts — same product, different surface.
1.4 · User Scenario Analysis
Three entry points, one architecture.
While defining the value story, I interviewed potential users recruited from Hugging Face and LinkedIn — mainly engineers familiar with evaluating new models. From their stories, the product clearly needed to support multiple entry points, not a single search bar.
Entry · A
Exploration without a clear goal
"I'm just curious what's out there." The user is browsing, not committing — wants to graze categories, see what's trending, leave with vocabulary they didn't have before.
Needs: discoverability · low-stakes browsing · trending signals
Entry · B
Task-driven discovery
"I have a job to do — find me a tool for it." The user knows the outcome but not the category. Comes with a vague goal, leaves with a shortlist they can act on.
Needs: intent parsing · curated comparison · decision-ready info
Entry · C
Feature-specific queries
"I want a tool that does this exact thing." Already mid-evaluation — comparing against alternatives, sometimes looking for an API or specific capability.
Needs: precise filters · capability search · API-level depth
1.5 · Information Architecture
I translated these needs into an architecture that balanced flexibility with clarity — keeping AI-driven suggestions at the heart of every entry.
Exploration
Task-driven
Feature-specific
Entries
Intent Parser
Shared core
AI Tools
GPTs
APIs
Collections
Result surfaces
Principle 02 in action
Watch the system show its understanding.
Click a prompt below — Nexa parses your vague intent into structured dimensions you can validate or correct. The interface speaks back.
Live mock — click another prompt to see the system re-parse the intent.
II · Design Iteration · 2.1 Overall Layout & Searching
Twelve directions. One survived contact with reality.
Group A
Left ⇄ Right
AI assistance docked beside the result feed.
- 01Side-Docked AI Suggestions Panel
- 02Search-First + Conversational AI
- 03Search-Enhanced AI + Result Feed
Group B
Top ⇅ Bottom
Conversation flows vertically; results fall beneath the input.
- 04Vertical Stack: Conversational Search
- 05Search Bar with Inline History + Card Feed
- 06Guided Search: Category Prompts + Result
Group C
Full Screen
Edge-to-edge launchers — preset prompts as the entry surface.
- 07Preset Prompt Launcher
- 08Search Expectations & Category Prompts
- 09Three-Panels: Output + Contextual Cards
Group D
Other
Inline, hover, and density-first variants — and the one we shipped.
- 10Inline Search Cards Within Conversation
- 11Persistent Bottom-Right Hover Agent
- 12Card-Only Mode for Max Result DensityShipped
"Visually rich and conversational layouts improved engagement during exploration, but introduced unnecessary cognitive overhead for search-driven tasks — and were intentionally deprioritized in the final design."
2.2 · Information Card
How much of a tool can one card tell you about?
Once layout was settled, the next question was the atom of the result feed: the card itself. Each direction below makes a different bet about what users need first — visual, action, comparison, or context.
A · Minimal
Avatar-Led Summary
Lowest visual weight — works at scale, but every comparison demands a click.
B · Discovery
Visual-Rich
Best for browsing. Visuals carry the meaning, but result density takes a hit.
C · Action-First
Direct-Access
Fastest path to "use the tool." Removes everything that isn't the verb.
D · Editorial
Editorial Style
Reads like a recommendation. Higher trust, slower scan.
E · Comparison
Relevance-Driven
Surfaces match strength. Great for shortlists; heavy on cognitive load.
For: copywriting
F · Use-Case
Use-Case Driven
Frames a tool by the job it's hired for, not by category.
G · Final direction
Horizontal — info and visual, balanced
The shipped card. Holds enough imagery to differentiate at a glance, enough copy to be decision-ready, and enough density to keep the feed scannable.
Shipped2.3 · Final Design
We landed on a search-first design that surfaces decision-ready information while intentionally minimizing visual distraction.
At that point, the core search experience had reached a level of clarity and stability that no longer benefited from incremental UX tuning.
The design challenge expanded — beyond usability, toward long-term platform sustainability and growth. Supporting creators, encouraging contribution, and accommodating advanced usage beyond UI-based tools.
-
Phase 1
Search experience
-
Phase 2
Card & result density
-
Phase 3
Creators · Submissions · APIs
III · Build AI · Build Ecosystem
Three modules to turn aggregation into an ecosystem.
Each module is anchored by an explicit Why? — the constraint that justified building it, not just the feature itself.
3.1 · Public Profiles
Creator pages & collections
Why? As the platform grows, trust, attribution, and continuity become increasingly important — especially for expert creators contributing tools over time.
- Persistent identity across submissions
- Curated collections as a first-class object
- Attribution surfaced wherever a tool appears
3.2 · Submissions
Direct contribution model
Why? Nexa already aggregated AI tools from open sources, but creators remained disconnected from the platform itself. Without a direct contribution model, tools exist as isolated entries.
- Submission workflow with state management
- Verification + ownership claim
- Status visibility from draft → live
3.3 · API Packages
Composability for advanced users
Why? Fully built tools meet most discovery needs — but advanced users often require deeper control, composability, and integration into existing workflows.
- Workflow: Choose → Build → Generated API
- Downloadable, reusable packages
- Surfaces parity with UI tools, not buried below them
Reflection · Results & Impacts
The platform operates at meaningful scale — providing real-world constraints and validation context for design decisions.
Reach
Monthly visits with balanced desktop and mobile audience
5M+
Per month
Engagement
Average visit duration — sustained engagement, not bounce
3min+
Per visit
Depth
Pages per visit — multi-step exploration behavior
4.2
Pages / visit
From the analysis, over 60% of users accessed via mobile — so we improved mobile-end responsive design as a direct consequence.
Thoughts
"When to stop optimizing individual interactions and start designing for system-level sustainability."
One key reflection was recognizing the moment to shift focus. Establishing a search-first foundation clarified user intent and reduced cognitive load — but long-term value depended on enabling identity, contribution, and advanced usage beyond search.
By introducing public creator profiles, creator submissions, and API packages, the platform shifted from passive aggregation toward an ecosystem that connects creators, users, and expert workflows.
If continued, the next phase would focus on validating these platform extensions through creator engagement metrics, API adoption signals, and behavioral indicators of long-term retention — while carefully managing the trade-offs between simplicity, governance, and scalability.