I'm Paul, lead designer on ECHO.

A defense intelligence platform handling 42 billion records.

Early-stage team, so I owned everything end-to-end: brand, product, research, and architecture.

Today's visuals are fully rebuilt on synthetic data for confidentiality.

The door you just used is the product.

Red means investigate. Rules are for humans, because the agent already has them.

I'll keep it to about 15 minutes. Feel free to interrupt with questions as we go.

The core problem: 42 billion records across 200 sources.

Yet an analyst still spent three days answering a single query.

The data wasn't missing.

The human was acting as the manual processing engine.

My design challenge: shift that cognitive load out of the user's head and into the platform architecture.

Quick context on scope.

Founder plus ten, and I was the only designer. So brand and product both.

I ran the research. I did the heuristic audit myself.

I designed the operator screens and the system underneath them.

It's confidential work, so everything you're about to see is rebuilt to the shipped spec on synthetic data.

Nothing here is a real investigation.

Three days. Four tools. One answer. That's where I started.

This is the old Overview.

Unlabeled icons. A wall of equal-weight cards dumping raw scores.

Redrawn on synthetic data so I can show it. The live screens stay confidential.

This is the old Personnel screen.

Eleven equal tabs. Score thirty-eight. Empty risk factors.

This is the old Identity screen.

Ten tabs. Slug titles. A card dump.

Seven surfaces failed this way. Average forty-seven. Redesigning it was the job.

One thing about the stakes, because it changes what good means here.

The cost of a miss isn't a service-level agreement. It's a life.

And Alex, my composite research persona, isn't judged on finding something.

He's judged on whether what he found holds up.

In a brief. In a court. Under a colleague's scrutiny.

I evaluated all 7 operator surfaces against one question.

Does this highlight the signal, or just add noise?

I actually started wrong.

Wasted time on high-level market positioning.

What actually cracked the case was dull, tactical work.

A strict Nielsen audit across every surface, to pinpoint where the interface was failing.

My method: audit a product the way an analyst audits a source.

Is this signal strong? Is it reproducible? Is it defensible?

So I ran a Nielsen heuristic evaluation across all seven operator surfaces.

Each one scored against a single question.

Does this make the consequential signal legible, or does it add noise?

One question, seven surfaces, so the scores are comparable.

That's what made the next slide useful.

The audit came back at 47 out of 100.

Dead-end charts. Unlabeled navigation forcing memory recall.

Request states vanishing between stages.

Different screens, identical root issue.

The interface was forcing the human to do the system's job.

Insight. Monochrome.

The fastest way to make one thing loud is to make everything else quiet.

My favorite design choice: the interface sits completely grayscale at rest.

Nothing here has earned attention yet.

Colour is reserved for meaning. Not decoration.

Teaching agents UI via color syntax and structure.

Same layout. Same components.

When a threat breaches a threshold, targeted colour activates.

Red signals an actionable alert. It never decorates.

Same red as Continue on the door.

From speaker: Raise, then Page. Phones Activate — or Seat. Resolved clears.

Every surface got the same treatment.

Name what the interface is making the analyst do. Then move it.

Threat trend gets 2x the grid: twice the area, and hue is rationed to tier.

Sidebar says the name: labels come back, and the case file stays pinned.

Every chart drills: no chart is a terminus. Every one drills through.

Pipeline stays on screen: stage, ownership, and an SLA countdown.

The analyst commands. The swarm does the reading.

Correlation, collection, the three-day grind — handed to agents.

The human role became judgment: what to pursue, what to cite, when to act.

Once automation handles data correlation, UX becomes a problem of trust architecture.

An analyst's work must survive high-stakes scrutiny.

So I made the model's reasoning explicit, dynamic, and bound to a signed chain of custody.

Every single insight is traceable back to its source.

Quick word on the rest of the terminal.

This is really seven surfaces, and I've only shown you one.

The feed is built for triage. Newest first, tier on every row, scannable at speed.

It's for deciding what deserves a look. Not for understanding it.

And the map is a genuinely different instrument.

Not the same data in a prettier wrapper.

Drag it. Routing reads as distance here.

That's a question a list cannot answer, no matter how well you sort it.

Which is the whole reason it earns its own surface.

To make this stick, I killed 1,000+ hardcoded colors.

Replaced them with tokens.

Then I wrote a build linter, so raw hex codes literally break the build.

It keeps the system tight.

And it gives AI coding agents a single source of truth.

My Design.md contract syncs directly with globals.css, Figma, and Cursor.

What you're seeing right now isn't static text.

It's reading live tokens from the stylesheet running this deck.

Same six tokens, three surfaces: the stylesheet, the AI agents, and Figma.

Tokens are how the design stays legible to the models. One source of truth.

This is the entire design token architecture.

Instead of organizing by raw color, I grouped tokens strictly by semantic role.

An achromatic baseline handles idle states, using just two text contrast levels.

Three dedicated signal colors communicate priority tiers.

Every value on this slide is rendering directly from the live stylesheet.

Same system holds from a 390 pixel phone up to a wall display.

It scales rather than just getting wider.

On the phone the nav becomes a drawer.

The feed keeps tier, threat and score. It drops only the supporting meta.

Eleven minutes.

The real-world result?

A federal operator cleared a target in 11 minutes instead of 3 days.

Every design decision directly restored attention back to the analyst.

I redefined the operational workflow.

Instead of forcing analysts to manually scan tables for hours,

the platform runs in the background.

It reaches out only when a trigger demands human judgment.

That's the core shift: from active manual surveillance to passive, event-driven interruption.

Two things I'd do differently, quickly.

First, sequence.

The upstream strategy work ran longer than the problem warranted.

A week of heuristic scoring outperformed all of it. I'd audit first now.

Second, the sample.

Both scores are audited, before and after.

But the operator-side numbers come from one federal customer, not a cohort. I'd want more.

What travels from this is three things.

Make the consequential signal legible. Show the reasoning, not just the answer.

And enforce the system in CI, not in a document.

You can test the system yourself via the live prototype link on screen.

That wraps the overview.

I'd love to open it up for any questions on the architecture, interactions, or design decisions.