Guide Instead of Teach
Agents shouldn’t need to rely on training or memory to complete an inspection. The system needed to deliver guidance at the moment it was needed - not before, and not after.
Case Study
HLEHandheld Lite Experience
Industry Context
Every vehicle handoff creates risk.
When a rental vehicle is picked up or returned, accurate documentation of its condition is essential. Without it, rental companies face damage disputes they cannot prove, missed damage they cannot charge for, and inspection inconsistencies they cannot explain.
The challenge compounds at scale. Different agents document the same vehicle differently. Under time pressure, photos get skipped, taken from the wrong angle, or captured with insufficient quality. When image quality drops, so does the reliability of the entire inspection process.
For AI-powered inspection systems, image quality is not a nice-to-have. It is the foundation that determines whether the system can do its job at all.
This challenge becomes especially difficult at locations without access to dedicated inspection infrastructure - the exact problem HLE was built to solve.
UVeye & The Coverage Gap
UVeye builds AI-powered vehicle inspection systems used by rental companies, fleet operators, and dealerships worldwide.
At sites where the hardware is installed, vehicles pass through an automated scanning tunnel that detects damage, tire issues, and safety anomalies within seconds - no manual effort required.
But not every site can install a tunnel.
Smaller branches, pop-up locations, and temporary operations face the same documentation requirements without access to the same infrastructure. They still need to record vehicle condition. They still need evidence when disputes arise. They still need consistency across every handoff.
HLE - Handheld Lite Experience - was built to close this gap. The goal was to extend UVeye's inspection capability to every site, using nothing more than the mobile device already in the agent's hand.
The Existing Experience
A mobile solution already existed.
But it relied entirely on the agent’s judgment - and that judgment was rarely consistent.
Users often struggled to understand
- what to capture
- which photos were mandatory
- where to stand
- how to frame the vehicle
- why photos failed validation.
As a result, photo quality was inconsistent and inspection data was frequently incomplete.
Images that couldn’t be reliably analyzed by the AI meant the system couldn’t deliver on its core promise.
The goal wasn’t to improve a screen. The goal was to redesign the entire capture experience and that’s exactly what this project set out to do.
Users & Design Challenge
The primary users were rental branch agents performing vehicle pickup and return inspections.
They were not professional photographers or vehicle inspectors. They were people working under time pressure - often between one customer and the next -
- in parking lots,
- in direct sunlight,
- in poorly lit garages,
- or in the rain.
What they needed was not capability.
They already had a phone. What they needed was clarity:
a system that made the right action obvious at every step, confirmed when they had done it correctly, and helped them recover quickly when they hadn’t.
This created the core design challenge:
How might we make photo capture simple enough for non-expert agents while ensuring the resulting images meet the quality requirements of an AI-powered inspection system?
Core Insight
Initially, the focus was on improving photo quality.
But field observations revealed something different.
Agents weren’t failing because they didn’t know how to use a camera. They were failing because at every step of the process, they didn’t know what was expected of them.
Where am I in this process?
Which part of the vehicle comes next?
Is the photo I just took good enough?
Why was it rejected? What do I do now?
This wasn’t a photography problem. It was an uncertainty problem.
That reframe changed everything. Once the challenge was understood as a guidance and clarity problem across an entire multi-step workflow - rather than a single-moment photo quality issue - the design response had to be fundamentally different.
Not a better camera interface. A system.
Design Principles
Before defining solutions, I established four principles grounded in what field observations revealed about how agents actually work.
Show Instead of Explain
Written instructions require interpretation. Visual guidance removes that step entirely. Seeing the correct framing is faster and more reliable than reading about it.
Recover Instead of Block
Mistakes are inevitable in real-world conditions. The system needed to help agents move forward - not leave them stuck in a failure loop with no clear path out.
Design for Reality
The experience had to work in busy parking lots, changing lighting conditions, and under genuine time pressure. Designing for ideal conditions would have produced a product that failed in the field.
A Structured Inspection Journey
The inspection required capturing 20 photos: 16 vehicle angles covering a full 360-degree view, and 4 tire photos - one per wheel. Together, these images create complete vehicle coverage and support accurate before-and-after damage comparison across the rental period.
20 photos
total inspection coverage
16 vehicle angles
covering a full 360-degree view
4 tire photos
one per wheel
Before HLE, this sequence was entirely unstructured. Agents had to remember what to capture, figure out positioning on their own, and judge whether each shot was acceptable - often while managing customers simultaneously.
The redesigned flow made the process explicit and navigable. A persistent progress panel shows all required shots as a visual checklist. Completed shots are marked as the agent moves forward. At any moment, the agent knows exactly where they are, what comes next, and how much remains.
The session was designed to be completable by a first-time user - without a trainer present, without prior experience, and without having to make a single judgment call about what the system needs.
Guided Capture Tutorial
Before starting an inspection, agents go through a short visual tutorial that introduces the capture process and sets quality expectations - without relying on prior training or documentation.
The tutorial covers common capture mistakes and provides clear examples of what inspection-ready images look like. By the time the first photo is taken, agents already know how to position the vehicle, how to frame each shot, and what to avoid.
Tutorial Overview
Tutorial Overview
Before the first inspection begins, agents complete a short visual walkthrough that establishes the mental model for the entire capture session. Rather than teaching photography, the tutorial aligns expectations, introduces the guidance system, and builds confidence before entering the live camera.
Learn
Ready to Capture
A key design decision:
Tutorial state is tied to the device, not the user.
In rental operations, multiple agents share a single device across shifts. Tying tutorial state to an individual login would mean one experienced agent could silently remove onboarding for every agent who follows.
Instead, dismissing the tutorial requires an explicit confirmation - a modal that makes the consequences clear: this affects all agents on this device, and the tutorial remains accessible at any time via "Watch Tutorial" during the session.
This one decision protected new agents from losing their onboarding without anyone realizing it.
The Core Solution: Adaptive Visual Guidance
The most important challenge was helping agents consistently capture inspection-ready images - in real time, across different vehicle types, locations, and lighting conditions.
To achieve this, I designed a visual overlay system that displays on the live camera feed for every capture step: a vehicle outline, a highlighted capture area, and framing guidance that shows the agent exactly where to stand and what to include in the frame.
A deliberate design decision:
Category-based overlays over model recognition.
Rather than attempting to identify each specific vehicle model - which would have required significant technical complexity and would have failed on unfamiliar models - the system uses eight vehicle categories:
- SUV
- Sedan
- Minivan
- Pickup Crew
- Passenger Van
- Hatchback (4-door)
- Hatchback (2-door)
- Sports
This created a scalable solution that works reliably across every vehicle an agent might encounter, without requiring the AI to know the exact make and model.
The overlay doesn’t need to be a perfect match.
It needs to make the correct framing intuitively obvious - and it does.
Guided Inspection Experience
Capturing inspection-ready images required more than a camera interface.
Since the AI could analyze photos after capture but couldn’t always identify in real time exactly why a photo failed, I chose not to rely on any single guidance mechanism. Instead, I designed a layered system where five complementary elements work together - each targeting a different moment of potential failure across the session.
A structured capture sequence removes the need for agents to remember what comes next. Visual progress tracking ensures they always know where they are in the process. Vehicle-specific overlays guide framing and positioning in the moment of capture. Reference thumbnails show the exact angle and area required for each shot. And contextual tutorials deliver support exactly when it’s needed - not before, not after.
The overlay system itself answers three questions before the agent presses the shutter:
Am I standing in the right place?
Am I capturing the correct area?
Is my framing correct?
Five elements make this possible - the vehicle overlay, the angle reference, progress tracking, just-in-time guidance, and optional guide controls that allow experienced agents to reduce guidance once they no longer need it.
Together, these layers transform a complex inspection task into a guided step-by-step experience that works even when individual components have limitations - because guidance doesn’t depend on any single mechanism being perfect.
Turning Photography Into Alignment
Rather than asking agents to interpret instructions, the system visually demonstrates what a successful photo looks like - and makes their task a simple act of alignment, not judgment.
Before capture begins, an animation plays that transitions from the vehicle preview to the live overlay. This single animation removes one of the most consistent early-session failure points: agents who understood the concept in the tutorial but couldn't apply it when they first faced the live camera.
The agent's only task becomes matching what they see through the camera to the shape on the screen.
No guesswork.
No interpretation.
No prior knowledge required.
Visual Guidance System
Vehicle Overlay
Align users with the correct position and framing.
Angle Reference
Shows the exact angle and vehicle area required for capture.
Progress Tracking
Keeps users aware of completed and remaining inspection steps.
Just-In-Time Guidance
Provides just-in-time assistance without interrupting the workflow.
Optional Guide Controls
Allows experienced users to hide guidance when it is no longer needed.
Guidance at the Right Moment
Guidance is most effective when it arrives at the exact moment it’s needed - not buried in a manual, and not delivered so early that it’s forgotten by the time it’s relevant.
Throughout the inspection, three moments of contextual support were designed to reduce uncertainty without overwhelming the agent.
Before the first photo is taken, a preparation prompt reminds agents to close all doors and turn off headlights - simple steps that directly affect image quality and are easy to overlook under time pressure.
At the start of each new shot type, a reference card shows a successful example of that specific angle or area, with a brief instruction. Agents see exactly what they’re aiming for before they raise the camera.
Different Capture Angles
For tire shots specifically, an illustrated guide shows the correct shooting position - crouching at center wheel height - because this angle is non-obvious and consistently produced errors without explicit guidance.
Just-In-Time Guidance
Guidance is most effective when it arrives at the exact moment it's needed - not buried in a manual, and not delivered so early that it's forgotten.
Three moments of contextual support were designed to reduce uncertainty without overwhelming the agent.
Before the first photo, a preparation prompt reminds agents to close all doors and turn off headlights - simple steps that directly affect image quality and are easy to overlook under time pressure.
At the start of each new shot type, a reference card shows a successful example of that angle or area. Agents see exactly what they're aiming for before they raise the camera.
For tire shots specifically, an illustrated guide shows the correct shooting position - crouching at center wheel height - because this angle is non-obvious and consistently produced errors without explicit guidance.
Progressive Error Recovery
Rather than blocking agents after a failed capture, the system provides progressively stronger guidance while keeping the inspection moving forward.
The recovery flow was designed as a gradual escalation: first explaining the issue, then reinforcing the guidance, and finally allowing the agent to move forward when the system could no longer be certain.
01
Identify the issue
Users receive a clear explanation of the detected problem and a recommended correction.
02
Increase guidance
If the same issue persists, the system reinforces the recommendation and encourages another attempt.
03
Allow recovery
To prevent workflow interruptions, users can either retake the photo or continue with the current image.
That third option was a deliberate and considered decision.
After three failed attempts, the problem may not be the agent. It may be the environment, the lighting conditions, or the system itself. Blocking agents indefinitely when the AI cannot confirm what is wrong would create frustration, erode trust, and prevent agents from completing a legitimate inspection. Offering "Use Anyway" acknowledges this honestly - it respects the agent's judgment while flagging the submission for downstream review.
Collaboration & Cross-Functional Work
This project was not designed in isolation. The quality of the outcome depended directly on close collaboration across disciplines throughout the entire process.
Product Manager
Working with the Product Manager, I helped translate business and user requirements into a coherent design strategy. We made joint decisions about what to prioritize, how to frame the guidance philosophy, and where to draw the line between what the system could support and what would need to wait.
AI & Computer Vision Team
Working with the algorithm and computer vision team, I developed a precise understanding of what the system could actually detect, what failure types it could reliably identify, and what it genuinely couldn’t know. This was not background research - it was core design input. The error escalation logic, the feedback language, and the decision to offer “Use Anyway” at Attempt 3 were all shaped directly by what the AI could and could not confirm with confidence.
Engineering Team
Working with Engineering, I made sure every interaction state was specified without ambiguity - tutorial state logic, error escalation conditions, animation behavior - so that design decisions were resolved before they reached code.
Design decisions were technically grounded. Engineering decisions were UX-informed. The product shipped with fewer gaps because ambiguity was resolved in design, not in development.
Designing Around AI Constraints
One of the most significant design challenges was that the system could not always explain exactly why a photo failed.
The AI could often detect that an image was problematic - but it couldn’t always determine the precise reason with enough confidence to communicate it clearly. Lighting, framing, distance, obstruction: the failure signal existed, but the specific cause sometimes didn’t.
This constraint shaped several design decisions directly.
Rather than displaying error messages the system wasn’t confident in, I chose to show feedback only for failure types the AI could identify reliably. Rather than demanding more from the AI than it could deliver, I built more resilience into the guidance layers themselves - so the experience remained helpful even when the system’s confidence was low.
The result was a deliberate trade-off: some specificity was sacrificed in exchange for honesty.
An AI-powered product that confidently gives wrong explanations loses user trust faster than one that acknowledges the edges of what it knows. At Attempt 3, saying “we couldn’t confirm this photo” was not a failure of design. It was the most honest and useful thing the system could say.
Design Tradeoffs
Every significant decision in this project involved a real tradeoff between competing needs.
Guidance vs Simplicity
Guidance depth versus interface simplicity: more guidance reduces errors but adds visual complexity. The layered system resolved this by making most guidance optional or contextual - present when needed, removable when not.
Validation vs Continuity
Validation strictness versus operational continuity: enforcing quality gates too aggressively would block agents from completing legitimate inspections. The three-attempt escalation model with “Use Anyway” kept the workflow moving without abandoning quality as a goal.
AI Specificity vs Trust
AI specificity versus user trust: showing only what the system could confidently identify - rather than generating plausible-sounding explanations - preserved the integrity of the feedback and protected long-term trust in the product.
Every UX decision solved more than one problem - and every solution required a deliberate compromise.
Impact
The redesigned experience produced more consistent image capture across agents, sites, and vehicle types.
Agents could complete the full 20-photo inspection independently, without prior training.
New staff onboarded faster.
And for the first time, mobile inspection sites could deliver the same quality of documentation as sites with permanent infrastructure - which was the original goal of the product.
Final Reflection
And the simplest lesson: simplicity for the user requires the designer to carry the complexity. The agent should feel none of the weight of the system behind the screen. Absorbing that weight - so they don’t have to - is exactly what the design is for.
The most important design move in this project was reframing the problem before committing to a solution.
Arriving with one assumption - that agents needed help taking better photos - and being willing to replace it with what field observation actually showed - that agents needed uncertainty removed from every step - changed every decision that followed.
Designing around AI limitations, rather than against them, produced a more honest product. When the system cannot explain why something failed, the design should not invent an explanation. Building more resilience into the guidance layers meant the experience remained useful even on the days when the AI’s confidence was low.