I want to build a repeatable process, not a one-off experiment

v1 compressed a typical 5-day design sprint down to a 1 hour solo activity.

I ran 3 AI assistants simultaneously to research the problem space - Claude, ChatGPT, and Gemini. I evaluated outputs in parallel, feeding them back into each other for cross-checking and narrowing.

5 minutes was plenty of time to figure out the problem to solve. Illegal deep sea fishing enforcement is difficult and impactful: there's too much data to sift through (i.e. Satellite SAR) and packaging the evidence is arduous. That's a perfect opportunity for AI-powered software to solve.

What I thought would be distinct research & brainstorm phases ended up happening at the same time. As I asked questions about the problem, the LLMs generated users, workflows, solution ideas without my always asking.

I wanted to start with a structured deliverable to act as a high level definition. I prompted with a short template to guide the research.

Deliverable 1: Elevator Pitch

The Elevator Pitch

🌊 Save the Ocean by Turning Satellite Data into Enforceable Evidence

Every year, roughly 1 in 5 wild-caught fish is taken illegally — generating up to $36 billion in illicit profits while accelerating ecosystem collapse and eroding a key natural defense against global warming. The vessels responsible don't need sophisticated tools to hide. They simply turn off their tracking transponders and "go dark."

Satellites can often see ships at sea — but seeing isn't the same as proving wrongdoing. Enforcement depends on structured, defensible evidence.

We're building an AI-powered Maritime Evidence Engine that detects dark-vessel behavior using satellite and vessel-pattern analysis, then automatically transforms that data into enforcement-ready evidence packs. Instead of just surfacing alerts, the platform delivers explainable, auditable reports that regulators and seafood buyers can use to deny market access, escalate inspections, and block illegal catch before it enters the supply chain.

We're not just watching the ocean. We're enforcing its laws.

Skipped

Process gap: v1 had no step to define the buyer. Without it, user and workflow were built on an assumption.

Fact checking: 1 in 5 and $36B unverified. Proceeded at risk.

After aligning all three tools on the elevator pitch, I used the AI tools to identify the core user & hero workflow.

Deliverable 2: Problem

The Problem

Illegal fishing + the evidence gap

Satellites can detect dark vessels. But detection isn't enforcement. Authorities need structured, defensible evidence — AIS gap analysis, SAR cross-reference, zone violations, vessel history — assembled fast enough to act.

The platform must: Collapse the "glitch, buoy, or bust" question from a 20-minute manual investigation into a 90-second confident decision. Speed to confident decision — not just detection.

Skipped

Fact checking: 20 min to 90 sec unverified. Proceeded at risk.

Deliverable 3: User

The User

Maritime Intelligence Analyst

Analysts at ocean NGOs, fisheries monitoring centers, coast guard units, and supply chain compliance officers at seafood retailers.

Core pain: Making defensible enforcement decisions with data not built for legal action. Bad escalations damage credibility — so they under-escalate.

More Info

Data volume with no triage

Thousands of anomalies surface daily. Everything looks equally urgent — or equally ignorable.

Explainability gap

They can see something looks wrong but can't articulate why in terms a regulator will accept. "The AI flagged it" isn't defensible.

Jurisdiction confusion

A vessel going dark in international waters triggers different rules than one in an EEZ. They're manually cross-referencing legal frameworks mid-analysis.

Tool fragmentation

AIS in one platform. SAR imagery in another. Port records elsewhere. Assembling a coherent picture means copy-pasting between systems — invisible, unrepeatable, legally fragile.

False positive fatigue

Escalating a bad flag to authorities damages credibility. So they sit on findings, waiting for certainty that never arrives.

Skipped

Process gap: no buyer defined meant the user choice was a guess. Risk noted above.

Retroactive: Persona icon and synthesis were done post-sprint, not during.

Deliverable 4: Workflow

The Workflow Chosen — Prototype Target

The "First Look" — triage in 90 seconds

1

The Trigger

An amber flash on their map. A detection flagged for review.

2

The Context Check

AIS silence — a track that stops while the vessel is still moving.

3

The Correlation

Overlay SAR. If a metallic signature is moving where AIS isn't broadcasting — heart rate goes up.

4

The Known Offender Check

Has this vessel been flagged before? Repeat behavior changes everything.

5

The Decision

Escalate with a timestamped reason, or Dismiss with logged justification — protecting the analyst if the vessel resurfaces.

Skipped

Process gap: v1 had no step to validate workflow choice against the 80% use case. Proceeded at risk.

The most fun: 7 prototypes in 30 min — across 6 tools

Each tool had varying amounts of context based on who had access to above conversations, which helps get variety in prototypes.

I thought Make and Tweak/Test would be two distinct sprint phases, but because each tool needed time to "think" it made sense to tweak ones while I waited for others to finish.

No winner selected. There were pros/cons to each and it'd be best to validate with users and more research.

Deliverable 5: Prototypes

7 Prototypes · 6 Tools

Claude · VS Code

Map + evidence focus. No alert list.

Claude · Browser v1

Dark Fleet command center. Tiny, dim fonts - poor readability.

Claude · Browser v2

Accessibility iteration, still hard to read.

Figma Make

Easiest to read. Confusing map & collapsible lists are a pain for this context.

Gemini

Triton Command had more approachable colors.

Lovable

Project Triton had the most cinematic sci-fi look but used biasing nationality signals (see below analysis).

Claude · Desktop

Dispatcher triage view had an interesting action bar at bottom.

Skipped

Human Judgement Corrections — too many to fix mid-sprint. See below.

AI moves fast. It also fills gaps with whatever conventions it knows. My job was to notice when those conventions were wrong for this context.

What I evaluated across the prototypes

Information hierarchy, accessibility, tone, and example data. All had pros and cons, so there's no clear "winner". Figma Make was the friendliest. Lovable was the most cinematic.

Tone: What AI defaulted to

Three of seven prototypes had targeting reticles on the map, which I hadn't asked for. The dark theme was partly my own doing (starting prompts), but the military surveillance conventions crept in on their own.

Example Data: What AI got wrong

Flags and real vessel names introduced political bias I wasn't comfortable with.

I caught it mid-sprint but didn't have time to resolve it all. Post-sprint I fixed the below example data. I also added tone and example data review to my v2 design sprint process.

Learnings from v1 of my Design Sprint Process

Design Sprint Process v1

Design Sprint Process v2

Not every next step is a full sprint. This phase validates before building further — and may loop before moving on.