Leo Chen

Reading the Future of Space, on One Screen

Company


Amazon Leo is a topic-intelligence dashboard for space sustainability. It reads regulatory filings, academic papers and press coverage, scores every topic on how widely it is being discussed, and predicts which ones are about to matter. I was the sole designer on a team of sixteen engineers and owned 100% of the design.

Role

UI/UX Designer (Contract via App Dev Club at University of Maryland)


Timeline

January 2026 - May 2026


Tools

Figma, React, Tailwind, Claude Figma MCP, Google Stitch, & other LLMs


TLDR


I designed for an intelligence dashboard that helps predict trends for space-policy analysts on space sustainability topics.


Before, most analysts had to read through every filing, paper, and article to find out. This would take days which is usually too late. As the sole UI/UX Designer with a team of 16 engineers, I turned a seven-layer ML pipeline into one screen an analyst can read in seconds. Key accomplishments included:

01 Information Design

Cut the five data spheres down to the three analysts actually use (research, press, and policy), and made every score traceable to its sources.

02 Stakeholder Alignment

Turned an idea Amazon's space sustainability engineer had carried for years into a buildable product.

03 Decision Support

Researched many other products that help condense trends. Replaced reading every filing, paper, and article by hand with one screen that flags which topics are about to reach regulation.

IMPACT


100%

of the front end and visual system was designed by me

7 —> 1

pipeline layers turned into one scannable screen

Greenlight

selected by Amazon's engineer to continue development for next year, with a conversation about using it as an internal analyst tool

"I've been thinking about this idea for years. The App Dev team turned my vision of the pipeline and made it way better than I'd imagined."

Yash C. - Space Sustainability Engineer (Regulatory), Amazon Leo

CONTEXT


A space-policy analyst, not a general audience


The reader is one person with one job: know what the field cares about before it becomes obvious. Debris, light pollution, collision avoidance, and reentry are argued over by regulators, academics, industry, press, and the public at once.

Five spheres feed the scoring model. The dashboard surfaces three, because an analyst reasons about research, press, and regulation.

INITIAL PROBLEM


All the information is already public yet nobody can read it.


Every document an analyst needs is already open to the world. Assembling it is manual, so the answer always arrives late.


Three core issues emerged:

Fragmented across spheres

The evidence is scattered across filings, papers, and press with no shared vocabulary.

Manual to track

Trends surface only after someone reads everything, after the moment to act.

Hype reads as importance

Loud is not the same as consequential, and nothing separates them.

RESEARCH


Key Insights


There was no analyst panel, so I talked to the professionals. I ran nine review sessions with Amazon’s space sustainability engineers, had 4 coffee chats with analysts in my network, and presented to the wider Amazon team. Three findings made me continuously iterate on the design:


  • Analysts think per sphere, not in aggregate. High globally but low in policy means the concern hasn't reached regulation yet.


  • Prediction beats popularity. The forecast, not the current volume, is what makes the tool worth opening.


  • Provenance is non-negotiable. An analyst won't cite a number they can't trace back to its documents.


RESEARCH


Key Insights


There was no analyst panel, so I talked to the professionals. After nine review sessions with Amazon's space sustainability engineers, 4 coffee chats with analysts in my connection, plus a presentation to the wider Amazon team. Three findings made me continuously iterate on the design:


  • Analysts think per sphere, not in aggregate. High globally but low in policy means the concern hasn't reached regulation yet.


  • Prediction beats popularity. The forecast, not the current volume, is what makes the tool worth opening.


  • Provenance is non-negotiable. An analyst won't cite a number they can't trace back to its documents.


GOAL


Create something to sift through noise and understand what the public actually cares about


With these insights, the design focused on what an analyst can safely ignore.


  • Show the gap between spheres, not one blended score

  • Put the forecast next to the history

  • Trace every number back to its sources


IDEATION


Three ways to show what matters


The reps asked for a word cloud, then leaned toward cinematic visuals. Each concept answered a different question, and only one answered the analyst's.


During ideation, I explored three approaches to showing what matters:


Using different LLMs, I looked at common approaches to designing an information focused topic explorer.

  • Word cloud (rejected): flattens everything except size

  • Topic explorer (rejected): shows how topics relate, not how much attention each carries

  • AI cinematic (rejected): spectacle over legibility

  • Orbital map (selected): size, position, and color carry three dimensions and stay readable

During ideation, I explored four approaches to showing what matters


Using different LLMs, I looked at common approaches to designing an information focused topic explorer.

  • Word cloud (rejected): flattens everything except size

  • Topic explorer (rejected): shows how topics relate, not how much attention each carries

  • AI cinematic (rejected): spectacle over legibility

  • Orbital map (selected): size, position, and color carry three dimensions and stay readable

Rejected - News Topic Explorer

Rejected - AI cinematic

Selected - Orbital Map

CONSTRAINTS


Working with Limitations in Time and Connections


I had no real access to working space analysts, so the review loop with Amazon's engineer stood in for usability testing while I interviewed a few junior analyst.


Once front end work started, some calls moved to engineering: the sphere palette changed, and the "Prevalence" rename didn't make it into the build. I'd rather show that than crop it out.

ITERATIONS & OUTCOME


The Full Story from Wireframe to Hand-off and Finally, Revisiting Passion


3 rounds of review with Amazon's engineers. Each round's notes added an iteration.

ROUND 1 - FIRST STRUCTURE

Feedback on the wireframe

Feedback on the wireframe

  • Trends will be interactive

  • Need a clear indicator for prediction in the timeline

  • Focus on the central hub map design

Size is prevalence, position is relevance, color is sphere. No sentiment coloring, because it belongs to keyphrases, not topics.

ROUND 2 - SECOND ITERATION

Feedback on round 1

  • Timeline could live in the map, where people are looking

  • No sphere selected should mean no score

  • Let analysts weight the spheres themselves

The timeline moved into the map, the ranking dropped to the top three, and the map got a date. A metrics dump became an argument: one score, the spheres behind it, a summary, and its sources.

ROUND 3 - HAND OFF

Feedback on round 2

  • Match the customer's terminology for the new specs

  • Hover shows users small info before clicking in

Prevalence and sentiment sit overlaid. Hovering and selecting now brings way more information without going deeper allowing users to decide if they want to explore further.

Prevalence and sentiment sit overlaid. Solid is history, dotted is prediction.

What was Shipped


The pipeline filled the map with hundreds of live topics. With five colors this would be unreadable; with three, it holds.

Where I'd take it next


After the semester, I kept going on my own even after the contract ended: the vertical axis now encodes time, so a topic's drift is visible without touching a control.

NEXT STEPS


Retrospectives


Restraint counted as much as addition

The hard part of a seven-layer pipeline is deciding what an analyst can ignore.


A stakeholder loop can't stand in for accessibility

Sphere is encoded by color alone, and at hundreds of bubbles color does all the work. Next time: shape and label before color, and a color-blind pass.


Design intent needs a seat in the build

The palette and rename changed after hand-off. Staying in the room through frontend would have kept them.

Next Steps

  • A sources explorer and a trends tab

  • A two-topic comparison mode

  • An accessibility audit before it reaches real analysts

Amazon Leo badge
Finish to collect
010203040506

Home

Vex Case Study

Home

Playground :)

Designed with intention and lots of R&B