A controlled experiment in AI-augmented UX design. Full discovery-to-handoff programme — 12 personas, 6 audience segments, research-led IA, and a complete design system — in 4 weeks. AI cut the timeline. It didn't cut the thinking.
Role: Solo UX Lead — research, strategy, IA, design, design system
Timeline: 4 weeks
Type: Personal project / controlled methodology study
Status: Research + Design Complete
StarJammer is a satellite tracking and space operations dashboard. It's not a client engagement. That's the point. I needed a contained environment to run a full UX programme and test a specific hypothesis: that AI can accelerate UX work without displacing the thinking that makes UX work actually good.
The claim that AI is replacing designers is everywhere. So is the opposite claim — that AI is just a tool and nothing has changed. Both are wrong. The real question is more specific: where in the process does AI add genuine value, and where does a human in the loop remain essential?
StarJammer was built to answer that question with evidence, not opinion.
The hypothesis: AI as workflow accelerant, not autonomous agent. Every strategic decision, research interpretation, and quality gate stays human. AI compresses the distance between thinking and artifact. It doesn't replace the thinking.
The result: The hypothesis proved out. The core strategic insight — that StarJammer needed persona-aware UI modes instead of a single expert interface — came directly from the research data. Not from a model.
Ran all discovery and definition phases: stakeholder framing, competitor analysis, heuristic evaluation against Nielsen's 10 principles, qualitative data coding, persona synthesis. No AI shortcuts on the reasoning.
Designed the IA and audience mode architecture from research findings. The mode-switching solution didn't come from AI ideation — it came from 12 personas with radically incompatible interface needs.
Produced all UX deliverables: 14-screen wireframe set, high-fidelity UI across 6 interface modes, and the Yamato 6.1 design system — tokens, components, WCAG 2.2 AA across light and dark themes.
StarJammer's dashboard delivered serious technical capability. It also assumed a single user type. A heuristic evaluation and persona analysis surfaced five interface failures — each with cross-segment impact, none of them edge cases.
Every new user hit a wall. The default interface dropped them into a fully-loaded expert dashboard with no guided entry. 8 of 12 personas experienced immediate, often terminal friction at first launch.
ISS passes, close approaches, satellite discoveries — none of it shareable. Six personas independently named this as their highest unmet need. A structural ceiling on growth sitting in plain sight.
The telemetry panel was live-only. Dr. Kenji (JAXA researcher persona) needed to examine eccentricity trends during a geomagnetic storm. That view didn't exist. Researchers had no path to the data they needed.
The AI Mission Assistant was a corner widget with a generic prompt. Seven personas could derive real value from contextual AI guidance — but nothing connected them to it or adapted it to their task.
The root cause across all five findings: One density, one language register, one feature hierarchy — for a radically diverse user base. Light/Dark toggle wasn't the answer. A mode-based architecture was.
Expert telemetry vocabulary with no plain-language equivalents. Casual users and educators had no way to interpret what they were seeing.
Mode-specific features were buried. Users couldn't discover capabilities relevant to their context without already knowing where to look.
No accelerators for expert users. No simplification paths for casual ones. One interface tried to serve everyone and fully served no one.
I ran structured discovery and definition through four research phases. AI was used in synthesis and documentation — not in the analytical work itself. Every model output was reviewed, revised, and approved by me before it influenced any design decision.
Central Research Question: "What interface, content, and interaction design changes will make StarJammer meaningfully usable across its full spectrum of users — without compromising depth for expert operators?"
That question has two competing requirements embedded in it. Resolving the tension between them — broad accessibility without expert compromise — was the design problem.
The mode-switching architecture wasn't a design principle I brought to the project. It was the only solution that fit the data. Twelve personas with radically different mental models, primary tasks, and feature priorities — all on one platform. One interface couldn't resolve that. Six modes could.
STEM Educator · Ghana
Needs a projectable classroom UI with simplified telemetry and a guided "Tonight's Sky" mode. Designed for a classroom wall, not a single screen.
Ham Radio Operator · UK
Needs precise orbital timing, pass prediction with azimuth/elevation, and Doppler shift calculations. High data density is a feature, not a problem.
Aerospace Researcher · JAXA
Needs historical telemetry trend views, API-level data access, and research export. The feature the platform was missing entirely.
| Segment | Representative Personas | Primary Need | UI Mode |
|---|---|---|---|
| Space Enthusiasts | Sofia, Kowalski Family | Discovery, wonder, sharing | Discovery Mode |
| Amateur Radio Ops | Marcus Webb, Yuki Tanaka | Pass timing, freq / Doppler data | Operator Mode |
| Scientists & Researchers | Dr. Kenji, Dr. Amara | Historical data, API access, export | Research Mode |
| Educators & Students | Ms. Chen, Sofia | Projectable, low-density, guided | Classroom Mode |
| General Public | Kowalski Family, James | Accessibility, large targets, plain language | Family Mode |
| Defence & Government | Agent Torres, Director Park | SSA overlays, classification, maximum density | Pro / Operator Mode |
The honest account: AI cut timeline on synthesis, iteration, and documentation. It didn't cut the thinking. Here's the actual breakdown — what stayed human, what AI accelerated, and what the time savings looked like.
| Phase | Who Did the Work | What That Looked Like |
|---|---|---|
| Research Design | Human | Research questions, methodology selection, scope definition — entirely mine. No AI involvement in how the study was structured. |
| Stakeholder Interviews | Human | Interview design, facilitation, and initial analysis. AI had no role here. |
| Qualitative Coding | Human | Inductive and deductive coding across the interface evaluation. The analytical interpretation — what the patterns meant — was my work. |
| Strategic Insight | Human | The mode-based architecture came from the research data. 12 personas with incompatible primary tasks made a single-interface solution untenable. That conclusion required understanding what the data meant — not pattern matching. |
| Research Synthesis | AI-Accelerated | After coding manually, I used Claude to accelerate affinity cluster labelling and theme naming. Every output reviewed, edited, approved before it influenced decisions. Estimated time saving: ~40% on synthesis documentation. |
| Concept Generation | AI-Accelerated | Used AI to generate divergent concepts for mode-switching patterns, onboarding structures, and the share layer model. I selected, combined, and refined. Nothing shipped directly from AI output. |
| UX Writing | AI-Accelerated | Prompted AI for UI copy — button labels, onboarding tooltips, contextual help text, AI Mission Assistant prompts — then edited for voice, clarity, and persona fit. Estimated time saving: ~60% on copy drafting. |
| Design Iteration | AI-Accelerated | Used Claude to review wireframe logic and flag cognitive load issues. Used AI to accelerate documentation — annotations, spec writeups, handoff notes. All output under my creative direction. |
The principle that held across the whole programme: Every AI touchpoint reduced clock time. None reduced the reasoning required from me as the designer. Synthesis, strategic decisions, and quality gates stayed human. The methodology stayed clean.
A 4-week timeline for a full discovery-to-handoff programme — 12 personas, competitive analysis, heuristic evaluation, qualitative coding, IA design, 14-screen wireframe set, high-fidelity UI across 6 modes, and a complete design system — would normally require 8 to 12 weeks of solo effort. AI compressed that timeline without changing the quality gates or removing the analytical work.
That's the case for AI as accelerant. Not as replacement. The outputs are only trustworthy because the thinking behind them was human.
The research finding was clear. Twelve personas across six segments didn't have different experience levels — they had fundamentally different primary tasks. The architecture had to reflect that. Six UI modes, each calibrated to a specific context, expertise level, and task set.
Casual users, first-timers, the Kowalski family. Simplified sky view, guided ISS pass alerts, shareable Story Cards for every high-emotion moment.
Ham radio operators. Pass prediction with azimuth/elevation, Doppler shift data, frequency window overlays. Everything Marcus needs, nothing he doesn't.
Scientists and researchers. Historical telemetry trend views, export, API access, annotation tools. The mode that filled the structural gap for Dr. Kenji.
Educators. Full-screen projectable layout, simplified data vocabulary, guided tour mode. Designed for a classroom wall, not a personal screen.
General public. Large touch targets (48px+), reduced information density, plain-language labels throughout. Accessibility as the default, not a setting.
Government and defence operators. SSA data overlays, classification tagging, maximum information density, secure context-switching controls.
Why modes, not progressive disclosure? Progressive disclosure still assumes a single feature hierarchy. Different segments have entirely different primary tasks — not just different experience levels. A farmer and a JAXA researcher both need depth. It's completely different depth. Progressive disclosure couldn't resolve that. Modes could.
Where does the AI Mission Assistant live? It's contextual, not a corner widget. In Research Mode it surfaces relevant telemetry summaries. In Discovery Mode it answers "What am I looking at?" In Operator Mode it feeds pass timing alerts. One AI feature — six context-aware expressions. The widget problem wasn't the widget. It was the lack of context.
How does the share layer work? Story Cards — auto-generated, shareable panels triggered by high-emotion moments (ISS passes, close approaches, milestones). Each card includes a map view, a data snapshot, and a one-tap share action. Designed to be the mechanism of organic acquisition: convert wonder into reach at zero marginal cost.
How does mode-switching work for users who don't know what mode they need? Onboarding routes users to a recommended mode based on a three-question context prompt. The mode can be changed at any time from the nav shell. Users aren't locked — they're guided.
Yamato 6.1 is the design system underlying every StarJammer mode. Built on atomic design principles, it delivers a unified visual language while giving each mode room to express its distinct information density and register. Every component was designed for WCAG 2.2 AA compliance from the start — not retrofitted.
Semantic tokens for surface, text, border, and accent — applied consistently across both light and dark themes. Mode-specific accent overrides handle the visual distinction between Discovery, Research, and Defence contexts.
Display, body, and mono type scales defined by viewport size and mode density. Family Mode uses a larger base size and higher line-height throughout. Research Mode supports data-dense mono rendering for telemetry values.
Three density tiers — Compact, Standard, Comfortable — map directly to mode contexts. Pro/Defence runs on Compact. Family Mode runs on Comfortable. Mode switching triggers a density change, not just a content change.
Duration and easing tokens defined for all interactive states. Prefers-reduced-motion support baked in at the token level — not handled as a special case per component.
The programme produced a complete, handoff-ready design package in 4 weeks. The AI integration didn't cut corners — it compressed timelines on the work that didn't require human judgment while leaving the work that did entirely in human hands.
The sharing layer targets viral discovery directly. Every unshareable wonder moment was a missed acquisition event. Story Cards convert high-emotion moments into organic growth at zero marginal cost.
Persona-aware onboarding, mode switching, and the elimination of the onboarding chasm reduce early churn across all 6 segments — particularly the 8 personas who hit terminal friction at first launch.
Classroom Mode and Family Mode open StarJammer to the education and consumer markets — segments the expert interface systematically excluded. New modes are new addressable market.
No direct competitor — Heavens-Above, N2YO, Stellarium — offers AI-assisted satellite operations with persona-aware UI modes. This is the category-defining gap the research surfaced.
The thesis proved out. AI as accelerant, not replacement. The core insight — persona-aware modes over a single expert interface — came from the data. The qualitative coding, the saturation testing, the strategic interpretation: all human. AI cut the time from research artifact to synthesis document, from concept to spec, from wireframe logic check to annotated handoff. It didn't cut the reasoning.
The replicable principle: AI integration works when it reduces the distance between a human decision and a human artifact. It breaks down when it's asked to make the decision itself. Keep the quality gates human. Use AI to move faster between them.