DreamCollege AI

Overview

Starting with a thank you

I appreciate you taking the time to review this case study. I chose this example because it demonstrates three things: shipping fast on a 0→1, high-visibility product; identifying and optimizing a key friction issue using real data; and influencing cross-functional roadmap decisions through user research.

Headquarters

Frisco, Texas

Founded

2023

Industry

SaaS · EdTech

Company Size

11–50

Skills

UX Research · Systems Thinking · Prototyping · IA · Usability Testing · Localisation

Platform

Web · Mobile · Conversational AI

Problem

A critical trust deficit in a $591B market

When I joined DreamCollege AI at the ideation phase, the college admissions landscape was facing a critical trust deficit. Students and families were making life-changing decisions worth hundreds of thousands of dollars — yet they were hesitant to trust AI with guidance that could shape their futures.

Core Challenge: How might we design an AI-powered college counseling platform that feels as trustworthy, personal, and reliable as working with a human counselor — while serving users at scale?

High stakes: College decisions represent $591.1B in the US education market in 2025
Trust barrier: 70% of prospective users expressed skepticism about AI for major life decisions
Market gap: Traditional counselors serve 1–2 dozen students; we needed to serve thousands
Timing pressure: 4-month runway with constrained budget and talent pool

Key problem areas identified in discovery

Conversational AI as the driver of all four outcome areas

Research

End-to-end design ownership — Research → IA → UX → UI → Testing

I led end-to-end design from 0→1, translating ambiguous AI-heavy business and user requirements into clear, user-centered experiences across: conversational AI interfaces, multi-step onboarding, payment flows, and core product features.

End-to-End Design

Led 0→1 design across conversational AI chatbots, hub-and-spoke onboarding, payment flows, subscription management, and all core product surfaces.

Design Foundation

Built a 48-component design system with tokens, naming conventions mirroring front-end frameworks, and Loom walkthroughs — cutting handoff time by 45%.

Drove Research

35+ student interviews, 28 usability sessions (Hotjar + Tobii eye-tracking), 12 stakeholder interviews. Iterative A/B testing on all critical conversion points.

Research scope — 75 total sessions across 4 methods

Research touchpoints breakdown (n=75)

Three trust barriers that shaped everything

Through 35+ student interviews and 12 parent/counselor sessions, three patterns kept surfacing. Naming them gave the whole team a shared framework for every design decision that followed.

"Black Box" Anxiety

Users felt uncomfortable when they couldn't see what the AI was thinking. "It's like asking for directions from someone wearing a blindfold." The fix: make AI reasoning visible at every step.

The Comparison Trap

Students weren't comparing us to other AI tools — they were comparing us to a "good counselor." The fix: personalization signals at every touchpoint.

Emotional ≠ Logical Stakes

Even when users logically understood AI capabilities, their emotional response was skepticism. The fix: progressive trust built through small wins, not big asks.

"It's like asking for directions from someone wearing a blindfold — even if they're right, I can't trust it."

— High school student, discovery interview

Process

Six phases across twelve weeks

Discovery and design ran in parallel due to the 4-month runway. Competitive analysis of 8 platforms, journey mapping across 6 personas, and heuristic evaluation of AI chatbots in education informed every decision.

Project timeline — 12 weeks, 6 phases

Discovery & Research

Weeks 1–2. 35+ student interviews, 12 parent/counselor interviews. Competitive audit of 8 platforms. Identified 3 trust barriers.

Problem Framing

Weeks 2–3. Established 3 design principles. Journey mapping across 6 personas. Aligned roadmap with CEO on Trust-First strategy.

IA & Journey Redesign

Weeks 3–4. Hub-and-spoke onboarding architecture. Information hierarchy redesign. Navigation structure validated with 8 users.

Visual System & DS

Weeks 4–6. 48-component design system. Design tokens for color, type, spacing. Weekly design-dev syncs to align implementation.

Conversational AI UX

Weeks 5–8. LLM response design, streaming states, trust layers. Drop-off reduced from 31% to 7%.

Testing & Iteration

Weeks 7–12. Hotjar heatmaps, Tobii eye-tracking, 12 A/B tests. Continuous iteration until all key metrics hit targets.

Design

Three north-star principles · 48 components · Zero design drift

Transparent by Default

Show users what's happening behind the scenes at every step. Visual cues for AI reasoning. If the AI can't explain it, it shouldn't say it.

Human-Centered AI

Make AI feel like an intelligent assistant, not a replacement for human judgment. Every interaction should feel like support, never surveillance.

Progressive Trust

Earn trust incrementally through small wins before asking for big commitments. Always give before you take. Users should feel confident before they commit.

Visual Design & Design System

Working closely with the CEO and branding stakeholders, I developed a visual language balancing approachability with credibility. A gradient-heavy palette (blues → purples) conveyed innovation while differentiating us from corporate competitors.

48 reusable components (buttons, forms, cards, modals, nav)
Design tokens for colors, typography, spacing, and shadows
Naming conventions mirroring front-end frameworks (e.g., btn-primary-lg)
Documentation for states, variations, and usage guidelines
Weekly design-dev syncs with shared Figma libraries

Result: Design-to-development handoff time dropped 45%, and developers could implement 80% of designs without clarification questions.

Color system — primary, accent, neutral, semantic

Full DreamCollege AI design system overview

Design system overview — typography, components, screens

Conversational AI — from 31% drop-off to 7%

One of the biggest UX challenges: keeping users engaged while the LLM generated responses (3–8 second delays). Initial testing showed 31% of users clicked away during loading, reporting they felt "ignored" or unsure if the system was working.

I designed a three-layer trust system: immediate acknowledgment (streaming animation), progressive disclosure (structured response format), and explainability (inline "why" callouts for every recommendation).

Three-layer AI transparency model

Drop-off −77% · Follow-up AI engagement +63%

Progressive profile building + transparent AI in production

Testing

Three decisions that changed the product — backed by data

Three critical design decisions with heatmaps

Decision 1: Profile-First Onboarding · Decision 2: Copy Profile from Sample · Decision 3: Collapsible Navigation

Continuous validation — heatmaps, eye-tracking, A/B tests

Hotjar Heatmaps revealed that users were clicking on non-interactive elements (icons, headings) expecting them to be buttons. I redesigned these as genuine interactive elements or removed them to reduce confusion.

Tobii Eye-Tracking (12 sessions) discovered users were skipping important context text styled too similarly to body copy. I introduced icon badges, colored callouts, and larger font sizes for key decision points.

12 iterative A/B testing rounds across critical conversion points
28 usability sessions using Hotjar + Tobii
12 stakeholder interviews to align on roadmap priorities
35+ user interviews across grades 9–12

Test

Original

Variant (Winner)

CTA Button Text
Onboarding save action

"Save and Continue" — 54% click-through

+8%
"Save My Progress" — 62% click-through

AI Response Styling
Chat bubble format

Plain text in bubbles — 3.2/5 trust rating

+0.6
Bullets + bold headers — 3.8/5 trust rating

College Card Layout
Browse/discovery view

Horizontal cards, small images — 23% click-through

+8%
Vertical cards, large images — 31% click-through

Insights

What the data surfaced that nobody expected

Stakeholder Management

The CEO wanted to prioritize monetization features (payment, upgrades) while research clearly showed we needed to nail core trust and usability first. I created a "Trust-First Roadmap" presentation showing user research quotes, conversion funnel data with a 47% drop-off before payment, and a proposed sequence: fix core experience → build trust → introduce premium features.

Outcome: CEO agreed to delay payment v2 by 2 weeks to focus on profile UX improvements. This resulted in 22% higher conversion when payment was eventually released.

The lesson: Stakeholder alignment isn't about winning arguments. It's about making the data so clear that the right decision becomes obvious. Frame around business outcomes, not design preferences.

Designing for AI is designing for the unknown

AI products introduce design challenges that traditional UX patterns don't solve: responses change as the model learns, user control must be preserved, and a user base spanning 30+ countries created unexpected localization complexity.

AI design challenges — surface vs. depth

Three things research found that nobody expected

Students Don't Read Instructions

Hotjar showed 94% of users skipped instructional text. We replaced text-heavy instructions with visual examples and progressive disclosure. Result: 41% improvement in form completion rates.

Parents Are Hidden Users

Eye-tracking revealed parents were frequently present — sometimes completing profiles themselves. We added a "Share with Parents" feature and rewrote copy for both audiences. Result: 19% increase in completed applications.

Anxiety-Driven Behavior

Students checked the platform 3–4× more than necessary. We built a "Your Journey" dashboard with "No changes since you last checked" status. Result: 28% reduction in unnecessary sessions.

Reflection

What I'd do differently — and what I'm proud of

What I'd Do Differently

Conduct usability tests in Week 2 rather than Week 5 — we could have caught navigation issues sooner and saved 2 weeks of iteration
Involve parents earlier as a distinct user group — their influence was underestimated until mid-project
Establish design tokens before the first production component — retrofitting is painful

What I'm Proud Of

Solving the trust problem — users feel confident making life-changing decisions because the AI earns trust through transparency, not authority
A design system architected to handle 10× growth without major refactoring — serving 150K today and built for more
Cross-functional leadership — bridging design, engineering, and business to translate user needs into measurable outcomes

Six lessons from building at the intersection of AI and trust

Trust is designed, not assumed

Users don't trust AI because it's capable — they trust it when the interface makes reasoning legible. Explainability is a design feature, not an engineering afterthought.

Reduce before you add

The biggest wins came from removing complexity. Simplification required more design judgment than building new features — and delivered more measurable impact.

Emotion is a UX signal

Students making life-defining decisions brought anxiety into every session. Designing for emotional state — not just task completion — was what set this apart.

Systems thinking multiplies impact

A well-built design system doesn't just save time — it creates organizational clarity. When everyone speaks the same design language, decisions get made faster and better.

Microcopy shapes confidence

At high-stakes moments, words do as much work as UI. Clear, reassuring copy around AI outputs reduced hesitation and helped users move forward with conviction.

Measure behavior, not opinions

Self-reported trust mattered, but completion rate, drop-off, and return usage told the real story. Instrumenting those signals made iteration faster and far more objective.

Design System & Collaboration

Weekly Design-Dev Sync

Demo new designs with interactive Figma prototypes
Review implementation against specs side-by-side
Discuss technical feasibility of upcoming features
Align on responsive behaviors and edge cases

Design Handoff Process

Figma files with organized layers, auto-layout, and interaction annotations
Assets exported at 1×, 2×, 3× with consistent naming
Spec document: user flows, copy, error states, success metrics
Loom video walkthrough (5–8 min per feature)

Conclusion

DreamCollege AI taught me that great design isn't about making things pretty — it's about building trust when stakes are high, simplifying complexity without losing depth, and creating systems that scale.

In 4 months, we went from a blank Figma file to a platform serving 150,000 students making one of the biggest decisions of their lives. The real achievement was earning their trust.

Glad we could cross paths.
Out of anywhere you could be, you're here.

150K users · 4.2/5 AI trust · 63% retention lift — in 4 months