VisualEyes: the first design assistant
Designing AI-driven design feedback, before AI tools for designers were a category.
VisualEyes turned attention models trained on real eye-tracking data into instant feedback on any design. I joined as the second hire and the sole designer, owning the end-to-end experience: research, IA, flows, and the shipped UI across the web app, the plugins, and the integrations.
- Product
- VisualEyes: AI-powered design feedback. Attention heatmaps, clarity scores, side-by-side benchmarks.
- Timeframe
- 2020 – 2021
- Role
- Product Designer · 2nd hire · sole designer
- Team
- 6-person startup; joined as 2nd hire and owned design end-to-end
01The bet
AI design feedback, three years early
In 2020, AI assistance for designers wasn't a category. Most teams shipped designs and waited weeks for data to tell them whether the designs worked. VisualEyes bet on bringing the feedback loop down to seconds.
From weeks to seconds
The bet was that designers would change their workflow if feedback was instant, embedded, and trustworthy.
Not 'AI-generated design'
We weren't generating layouts. We were measuring them: attention and clarity, not aesthetics. A different and harder pitch in 2020.
Where designers already work
If the feedback didn't live inside Figma / Sketch / XD, it would lose to opening another tab. Integration-first from day one.
02Process
From ideation to product
Five phases, overlapping by design: research and architecture seeded the MVP; UX, UI, and development ran in parallel; evaluation loops fed every next iteration.
03Research
Two audiences, one product
Personas and job stories anchored every screen: who's doing the work, and what they need to do faster.
Personas
UI Designer
ICs shipping screens on a deadline.
Validation in the moment, not after the fact. Clear, actionable signals.
Manager / Lead
Reviews work, makes calls, benchmarks against competitors.
Comparable scores, share-ready reports, an audit trail of what shipped and why.
Job stories
- 01
“I want to use the product I just purchased without leaving my tool of choice.”
- 02
“I want to organize predictions into projects I can revisit and share with the team.”
- 03
“I want to compare two designs and show the team which won, and why.”
- 04
“I want to manage my account, plan, and team access without filing a support ticket.”
04Architecture
The map of the product
A multi-level IA spanning auth, dashboard, projects, predictions, guides, subscription, and account, designed so the same patterns held across the web app, the plugins, and the API.
05Surfaces
What designers actually used
Four screens that carried the workload of the product. The problem each one solved and the response.
First-run guidance, not a wall of empty state
- Problem
- New users land on an empty product. The first prediction is a leap of faith.
- Design
- A home that opens with integrations, curated guides, and a single 'start your first prediction' CTA, so the next click is obvious.
Organize, revisit, share
- Problem
- Predictions piled up as one-offs, hard to revisit or compare. Teams needed structure.
- Design
- Projects as containers: group predictions, share with collaborators, and revisit them in context.
Compare two designs, show your work
- Problem
- Designers needed to argue for a direction, and managers needed evidence, fast.
- Design
- Side-by-side compare mode with shared scoring, plus quick add-prediction flow so a project grows during a review.
Three tiers, clear gating, no rug-pulls
- Problem
- Mixing a generous free tier with paid plans is a trust problem. Surprise downgrades break adoption.
- Design
- Free / Basic / Pro with explicit gating and self-serve plan changes; downgrades warn ahead with a grace period.
06The decision
Plugin-first, web app second
The hardest call: where does the product live in a designer's day.
A great standalone web app is easier to build. But designers spend their day inside Figma / Sketch / XD. Asking them to switch tools is asking them to lose.
Web app first
✕ Loses adoptionFaster to build, full UI control. Every prediction means a context switch.
Plugin first
✕ Loses velocityLives where work happens. Each tool is its own build, with constrained UI.
Both, in parallel
✓ ChosenPlugins for the daily loop. Web app for projects, sharing, and team work.
Tradeoff accepted
More surfaces to design. Tighter constraints inside each plugin. We accepted both to meet designers where they were.
Result
VisualEyes shipped across Figma, Sketch, Adobe XD, and Chrome, plus a web app for teams. The plugin became the daily loop; the web app became the report.
07Reflection
What shipping AI tools for designers in 2020 taught me
What landed
- The integration-first call. Plugin adoption beat web-only competitors that came later.
- Treating AI output as a design material, not a verdict. Heatmaps and scores worked better as inputs to a designer's judgment.
- Picking two opinionated metrics (attention, clarity) over a dashboard of every signal we could compute.
What didn't
- Underestimated how much trust takes: designers wanted to peek at the model before believing the result.
- Some flows were too web-app-shaped and didn't translate cleanly into the constraints of a plugin.
- The free tier was too generous, which softened the upgrade story more than it needed to.
What I'd carry forward
- Design for the workflow, not the screen: the difference between adoption and a great demo.
- Make AI feel like a colleague who's clear about its limits, not a judge handing down scores.
- Onboarding an AI product is mostly trust onboarding. Surface the model, not just the output.