Louis Wu, Design Engineer / HCI / UX
The vehicle image from the original Crossover Car concept

Crossover Car

AI-assisted
car buying
decisions

From an initial idea to a clearer buying brief, visible evidence and an informed next step.

View design response Browse prototype

Independent iteration Working prototype

Role
UX & AI interaction design
Functional prototyping
Project
COM 515 team foundation
Independent iteration, 2026
Methods
Journey mapping, information architecture, interaction design, prototyping and interface iteration
Focus
Preference elicitation
Used-car decision journey
Evidence and user control

Crossover Car

Crossover Car is an AI-assisted decision-support prototype for used-car buyers. It helps people clarify what matters, explore suitable vehicles, inspect source information and compare trade-offs.

The aim is not to choose a car for the user, but to help them understand their priorities and decide what to check next.

Context

The project began as a COM 515 team exploration. Building on that shared foundation, I independently reframed the experience around preference elicitation and developed the current interaction design and working prototype.

The early research and wireframes below document the collaborative phase; the current prototype reflects my subsequent individual iteration.

Design Challenge

Used-car buyers face scattered information, uncertain vehicle histories and priorities they may not yet be able to articulate. How could an AI-assisted experience help them clarify their needs, understand the evidence and stay in control of the decision?

User Journey Map

Wireframes

Onboarding
Early team onboarding sketch
AI Question Flow
Early team AI question flow sketch
Vehicle Listing Page
Early vehicle listing wireframe
Preference Selection
Early preference selection sketch
Homepage Layout
Early homepage wireframe
Overall Interface
Early interface wireframes arranged on a table
Goal-Oriented Buyer
Early goal-oriented buyer flow
Guided First-Time Buyer
Early guided first-time buyer flow

Qualitative Study

During the team phase, we explored how buyers approached used-car decisions and responded to the early AI-assisted concept. The study surfaced questions about clarity, trust and control that informed my later redesign.

Research Setup

The original team study involved four participants using a low-fidelity prototype. These findings are formative input, not validation of the current working prototype.

Participants
4
Method
Moderated usability interviews
Format
Think-aloud
Duration
10-15 minutes each
Prototype
Clickable low-fi prototype

Participant Snapshot

P1

Graduate student / used-car buyer / concerned about hidden maintenance costs

P2

International student / prospective first-time buyer / wanted to browse before logging in

P3

General tester / focused on AI tool clarity / wanted a human-support option

P4

Recent buyer / limited car knowledge / concerned about missing alternatives and source reliability

What We Tested

Homepage Entry

Start with AI vs. browse cars

Car List + Detail Flow

Listings, filters and vehicle information

AI Evaluation

AutoCheck labels, scores and explanations

Comparison

Side-by-side trade-offs and decision support

Key Findings

The early sessions pointed to a need for explanation and evidence, not just a recommendation.

01

Browse First, AI Second

Participants hesitated between the two entry points. Familiar browsing needed to remain available.

02

Scores Need Explanation

A score alone was not enough. Participants wanted to understand the factors, reasoning and sources.

03

AutoCheck Was Ambiguous

The label did not clearly communicate what had been checked or what was still unknown.

04

Comparison Supported Discussion

Side-by-side information gave participants a way to discuss the trade-offs between vehicles.

05

Trust Requires Evidence

Questions about reliability and missed alternatives made source visibility an important design direction.

From Research to Design Direction

I used these early findings to refine the journey: clarify preferences, explain recommendations, separate fit from condition, and support comparison.

Progressive preferencesVisible reasoningSource evidenceComparison support

Design Response

Building on the early study, I independently developed a working prototype around preference elicitation: helping buyers turn an initial idea into a brief they can confirm, then inspect evidence and compare trade-offs.

Try the guided prototype (opens in a new tab)

AI Advisor Vehicle evidence Compare

01 Clarify before recommending

Small choices make priorities explicit before the user confirms a buying brief.

02 Separate fit from condition

Show seller photos, disclosures and missing evidence separately from recommendation reasons.

03 Compare the trade-offs

Keep differences side by side so buyers can decide what to inspect, ask or rule out.

Curated walkthrough: one archived real listing and one fictional comparison. Not live inventory or an AI vehicle inspection.

THE DESIGN LOGIC / PREFERENCE ELICITATION

A first thought. A clearer direction.

AI helps clarify the meaning. The buyer makes the call.

THE FIRST MESSAGE

“Something for weekends. Maybe something fun?”

Set a boundary

What feels comfortable to spend?

Example limits in USD, before tax & fees. Not market price bands.

1 / 4 · Set your budget

Interactive illustration · Scripted AI responses
About this example

Illustrative choices, not a participant quote or a live model call. This demonstrates the intended preference-clarification interaction; the connected prototype supports a narrower set of preferences.

PREFERENCE ELICITATION / BEYOND THE FIRST MESSAGE

What you explore can start a better question.

Not every preference starts in a chat. Browsing can reveal a clue, but the buyer decides what it means.

Loading the browsing example.

In the full prototype: local storage · Last 7 days · Up to 100 events · No outside history.

Open the full prototype (opens in a new tab)
Where AI fits, and what is actually tracked

The embedded example runs only in memory and resets on reload. Its five fictional listings use illustrative attributes with the prototype's 40 / 35 / 25 budget, use and priority weights. Photos represent the models, not the exact vehicles. This prepared example changes two Home picks after confirmation; it is not live inventory, a market valuation or an AI vehicle inspection.

In the full prototype, browsing assistance is off until you enable it. After permission, it stores a vehicle ID, action and timestamp in this browser for views, saves and comparison choices, including removals. It does not calculate site-wide page views or click-through rates. Local rules suggest space or fuel efficiency only after activity on at least two different vehicles, including a save or comparison; one click is not enough. Nothing changes until you confirm. With a complete brief, Home ranks available sample cars and shows up to three; the same cars may remain if they still fit best.

These browsing suggestions are rule-based, not a cloud AI model. The design gives an AI-assisted conversation useful context; a separate model review would require a connected provider and explicit permission to send an activity summary. The hosted demo has no live model connected and does not run background cloud analysis.

Model photo credits

WHAT CHANGES

The difference is the connection.

FILTER-FIRST SEARCH

I know what I want.

CriteriaListings

GENERAL-PURPOSE LLM

Help me think it through.

ConversationAdvice

Clarifying and comparing are not unique to Crossover.

CROSSOVER

Carry my priorities forward.

BriefEvidenceDecision
WITHOUT AI

I translate my needs into filters.

WITH AI ASSISTANCE

I describe my needs and confirm the meaning.

What is working now, and what is still a design hypothesis?

Without a connected model, guided choices, filtering, evidence views and comparison still work. When connected, AI helps interpret free-text needs and proposes a supported use or priority for confirmation. It does not inspect vehicles or calculate the ranking.

The linked walkthrough is a prepared example, not a live AI session. Owner-review summaries are AI-assisted, source-checked snapshots. Reduced effort and greater confidence are design goals, not outcomes validated by the early study.

This comparison describes interaction patterns, not limits on every marketplace or chatbot. OpenAI documents intent clarification in ChatGPT and research and comparison capabilities.

From “Which car should I buy?”
to “I know why this option fits, and what I need to verify.”