Dell Research Dashboard | Dhruvi Gopani UX Case Study

Back to Works

✅ Shipping · Enterprise

Turning thousands of research reports into decisions executives could make in minutes.

An insights dashboard that turns scattered user research into decisions — the interface layer of a larger research-automation system at Dell. Grounded in interviews with one internal product team, where research too often stalled in slide decks and inboxes. The thinking and the trade-offs are here; validation in progress.

ROLE

Sole UX Designer

TOOLS

Figma · Gemini · Claude

TIMELINE

10 weeks

TYPE

Internal tooling · AI-assisted

Team specifics are kept confidential and sample data is AI-generated; examples shown are illustrative.

TL;DR

Problems

Research lived "in people's heads."
Insights pieced together from fragmented sources.
Big reports created information overload.
Findings lacked prioritization, consumed second-hand.

Solution

Designed and researched in parallel.
Made AI find and draft; kept humans deciding.
Surfaced summary-first, confidence shown.
One system, three doors — a view per role.

Results

An end-to-end high-fi system, presented to Dell.
A defensible line on where AI belongs.
Built to the target: minutes, not hours.
A concept the team could carry forward.

THE DIFFERENTIATOR

The hard part wasn't adding AI. It was knowing where to stop.

The brief was to turn research data into decisions. The lazy version is "add AI" and let it summarize everything. The real one was harder: figure out where AI earns trust — and where it can't have it.

The people I interviewed answered that for me. They already half-distrust AI — and they distrust it in a specific place. They'll happily let it find. They will not let it decide.

AI search through context is very useful

So the whole system rests on one line: AI retrieves and drafts. Humans confirm and decide. Every insight it surfaces links back to its source. Nothing gets to sound authoritative without being traceable.

💡

The dashboard never speaks for a researcher. It just helps them speak faster.

CONTEXT

It looked like a search problem. Until you asked how research actually moved.

Hundreds of studies sat across multiple repositories. The tooling existed. But finding anything meant tracking down whoever ran it and hoping it was still in their head.

Trends I notice stay in my head — they aren't written down anywhere

People tune out during 45-minute presentations

They just want a list of things to fix

So people improvised — asked friends, ran quick studies, built private piles of insight that died the moment they moved on. This was the state of things when I joined: a small brief, a tight timeline, and a lot I didn't know yet.

WORKING IN PARALLEL

I couldn't wait for research, so I designed toward it.

Scheduling six busy people took weeks, and I didn't have weeks to sit still. So I built a first set of mid-fi flows from the brief — not to ship, but to have something concrete to be wrong about the moment the interviews landed.

Two tracks at once: prototypes moving forward, research catching up to them. By the time I got people in a room, I wasn't asking abstract questions — I was pointing at real screens and watching my assumptions break. I worked alongside a researcher throughout, with weekly check-ins to pull data and clarify — and part of the craft was simply navigating a brutal calendar without wasting a slot.

SIX INTERVIEWS WITH ONE PRODUCT TEAM

I expected a findability problem. I found a people problem.

01 · Research moved by asking people, not searching a system.

Access meant someone's memory and availability. When they leave, the index leaves with them.

02 · Everyone worked on a different clock.

Specs this quarter, usage data next year, trends two years out. One flat repository fails all of them.

03 · Stakeholders wanted the summary first.

"Executive summaries help people decide if they care." Depth was for later, on demand.

04 · Interpretation was gatekept.

Even when research was findable, acting on it wasn't always allowed. Access was political, not only technical.

05 · AI was trusted to find, never to decide.

They'd let it search all day. They'd never let it write the conclusion.

💡

This was never a findability problem. Research lived in people — their memory, their authority, their attention. A search box wouldn't fix that.

Honest craft note: I couldn't touch confidential research, so I used ChatGPT, Gemini, and Claude to generate realistic mock study data — plausible titles, quotes, and stats — so copy and layout could be pressure-tested without exposing anything real. 🙂

Fig. 01. Left: the interview synthesis that reframed the brief. Right: the parallel-track sketches it reshaped.

PRINCIPLES

Four rules that kept every screen honest.

01 · Clarity over complexity. Reduce cognitive load before anything else.

02 · Progressive disclosure. Most important thing first, depth on demand.

03 · Action-oriented insights. Every finding points at a decision.

04 · Scalability. It has to hold up as the research pile keeps growing.

Show the confidence, don't hide it.

If AI is drafting and retrieving, the fastest way to lose trust is to make everything look equally certain. So I did the opposite of most AI products and put the doubt on the surface. Every study carries a confidence level and a Data Limitations panel — sample skew, age gaps, small-n caveats — right at the top, before the numbers.

A tool that hides its limitations is just a confident liar.

Summary first, everything later.

Hundreds of studies, four roles on four clocks. One wall of data served none of them. So the study view opens with the five things that matter — executive summary, priority insights, associated personas, actionable recommendations, and links to go deeper. Drill-down is one tap away, but nobody has to earn the overview.

A stakeholder shouldn't have to know the right question before the tool is useful to them.

One system, four doors.

Researchers, PMs, designers, and business stakeholders all needed this, and each wanted something different. Four bespoke tools would've been impossible to maintain. So: one core system, role-aware. Semantic search takes a plain phrase — say, "onboarding friction" — and returns the studies, personas, and insights tied to it, each tagged with method and confidence.

Same interface. Different ways in.

Design the upkeep out.

Every repository dies the same way: upkeep falls on researchers, competes with real research, gets deprioritized. And I'd heard it directly — they'd rather fire off a quick email than sit through 1:1 syncs. So I paired the dashboard with an AI co-pilot: drop in a report, get a standardized summary — finding, insight, recommendation, method, deep-dive link — in about two minutes instead of an afternoon. The library fills itself as a byproduct of work they already do.

If upkeep depends on goodwill, it dies. I designed so it didn't have to.

Fig. 02. Decision 01 in the UI — the doubt shown first, so people know how far to trust each insight.

Fig. 03–04. Left: one query, the whole trail behind it. Right: the two-minute summary that keeps the library alive.

HOW I USED AI

I ran my own process on the same rule I designed for.

The irony wasn't lost on me. So I held myself to the product's own line.

AI did the grunt work. Organizing insights, drafting copy, brainstorming dashboard structures, rapid concept exploration.

I kept the judgment. UX strategy, prioritization, information architecture, and every visual design decision.

AI accelerates. It doesn't get to decide.

WHAT IT DELIVERED

A system that let research move without a person attached to it.

This was a capstone built in collaboration with Dell, so the proof is in what reached the room, not a production metric.

An end-to-end high-fi system, presented to Dell. Study detail, searchable library, semantic search, persona linking, and the co-pilot — grounded end-to-end in the discovery.

A defensible POV on AI in research workflows. Retrieve and draft, never decide. Show confidence and limitations up front. Make the human the accountable step.

Designed to a clear target. Research understood in minutes, not hours — summary-first, prioritized, traceable.

A direction to carry forward. Predictive insights, more automation, expanded reporting.

WHAT STUCK

The problem wasn't findability. It was that research moved through people.

Once the system captured what lived in someone's head — showed how much to trust it, and respected who was allowed to act on it — the friction had somewhere to go.

Two things stuck. Starting the prototypes before the research was fully in wasn't the neat version of the story, but it's the one that shipped on time — designing toward research beats waiting for it. And AI is most useful the moment you decide where it stops. The restraint was the design.