Case study · Analytics platform
Meridian
One analytics platform for an entire university, built so a dean gets an answer in about two minutes instead of six.
The brief
One institution, four dashboards, five months
My role
- Lead UX designer, end to end
- Heuristic audit and secondary research
- Information architecture and user flows
- All eight screens and the data-viz system
Timeline
- 14 weeks
- 3 discovery
- 8 design
- 3 validation and handover
What they already had
They had the data. They could not read it.
- Four separate Power BI dashboards: undergraduate admissions, graduate admissions, research and HR.
- Four owners, four vocabularies, and no shared definition of a single metric.
- Leadership kept saying the same sentence: I have the numbers, I just cannot find the answer.
- The brief I was handed was make the dashboards nicer. That would have produced a prettier version of the same problem.
Baseline
I measured the old system before I touched it.
Six numbers from the audit. Every one of them became a target.
Heuristic scoring
Four dashboards, scored screen by screen.
Screen by screen, every issue rated 0 to 4 for severity and tied back to a real person and a real task rather than to a rule. Four findings did most of the damage.
Two passes per dashboard. The first was task-based, walking the real questions leadership had asked that month and noting where I stalled. The second was a straight heuristic sweep against Nielsen's ten, so nothing got excused just because I had learned my way around it by then. Every issue got a severity and the name of a person it actually cost, which is what made the four that mattered separate themselves from the forty that did not.
Filters reset the moment you navigated
Every dashboard put its controls somewhere different and forgot them on the way out. Rebuilding the same cohort three times a day is not a preference problem, it is a memory tax.
One number described nobody
Undergraduate and graduate figures were blended into averages that matched no population on campus. People were making real decisions on a number that did not exist.
Every view was a single year
Nothing on screen showed direction, so any question about a trend became a ticket to the analytics team. That is why leadership had data and no answers at the same time.
Everything was emphasised, so nothing was
Twelve equal-weight tiles per screen, and eleven pie charts across the four tools. Reading a comparison meant judging angles across separate charts, which people are measurably bad at.
Five of the findings became design rules.
A page of twelve equal weight tiles, where nothing is first. The fix leads with one headline KPI, then a trend, then the detail.
A pie of near equal slices is hard to compare. The rule: shares stay donuts, comparisons become sorted bars, time becomes lines.
Four disconnected pages with no links between them. One overview cockpit drills down into four modules. Nothing is a dead end.
A flat gray number, no prior year, no direction. The fix carries the value, a colored delta, an arrow, a sparkline, and vs last year.
A dense page spends its first 90 seconds orienting. The answer first page spends that time for the user, so it reads instead.
What the scores add up to
The problem, in one sentence
University leaders cannot answer how the institution is doing, because the answer is split across four tools that each define their numbers differently.
- Every question means opening four dashboards and reconciling them by hand.
- The same word means different things in each one, so the reconciling is guesswork.
- Decisions get made on the number someone remembers rather than the number that is true.
The sector pattern
What peer universities do about this, and where Meridian broke from the pattern
Dashboard sprawl is the sector’s default state, not this university’s special failure. The common answer is to buy more dashboards: a BI license per department, each with its own definitions, refresh cycles and owner. EDUCAUSE has ranked data and analytics governance among higher ed’s top technology issues for years, and the pattern behind it is always the same. Institutional research becomes a report factory, every leadership question becomes a ticket, and cabinet meetings open with an argument about whose number is right, because “enrolled student” means three different things in three different files.
Meridian broke from that pattern in three ways. One platform instead of one per department. One definition per metric, signed off by the analysts who own the data, so the argument about whose number is right ends before the meeting starts. And answer-first screens, so the platform replaces the report queue for routine questions instead of adding to it.
The three tests
Every screen had to pass all three.
- Trusted. Every number traces to the Common Data Set and IPEDS. One definition, one owner, no exceptions.
- Readable. The headline first, the detail on demand. Two minutes has to be enough to get an answer.
- Segmented. Nothing blends across levels. Undergraduate and graduate never share a number.
The reading problem
One blended number was quietly lying on every screen.
- Meridian's graduate applicant pool is 48% international. The undergraduate pool is 12%.
- The old dashboards blended the two into a single figure of about 22%.
- That 22% describes an applicant who does not exist anywhere. It is like averaging your oven and your freezer and calling it room temperature.
- So it became a hard rule: population mixes never blend across levels.
What leadership asked for
The questions they could not answer
- The dean. A couple of minutes between meetings. Wants the headline and whether it is moving the right way.
- The analyst. Happy to spend an hour pulling a cohort apart, as long as the numbers hold up.
- Both. One screen has to serve them without compromising for either.
Delivered in phase one
- Overview cockpit. All four departments at a glance, one question answered.
- Four department tabs sharing one navigation and one metric dictionary.
- Multi-year trend by default on every view, with year-on-year change on every number.
- Ask Meridian, a plain-language panel that explains the screen you are on.
Who reads this
Built around reading speed, not job titles.
The useful split was not one job title against another. It was how long each person has, and what they are willing to do to get an answer.
Built from interview transcripts rather than from the org chart. I coded every session for two things: how long the person had before they needed an answer, and what they did when the dashboard did not give them one. Those two axes separated people far more cleanly than their titles did, which is why a dean and a department head ended up in the same group and two people from the same office did not.
Senior leadership
Wants the whole institution in one glance: the headline across all four domains in ninety seconds, multi-year trends, and forecasts flagged as forecasts.
Enrollment management
The most demanding audience. Lives in the funnel: applied through to enrolled plus summer melt, yield by segment and source market, deposits tracked all summer.
HR leaders
Own their own data. Headcount, turnover, and academic mix, with faculty and staff kept apart and workforce composition presented carefully.
Feature flow
The route from landing to decision.
Writing the path down first is what exposed how much of the old experience was spent orienting rather than reading.
Open Meridian
Land on the overview cockpit, no filtering needed.
Read the headline
Up or down, against last year and plan, anything on fire.
Open a module
Undergraduate, Graduate, Research, or HR, segmented by level.
Filter to a cohort
Program, geography, test policy, term. The funnel updates.
Act or export
Trigger outreach, flag a risk, or export for the board deck.
Glance, trust, leave. The cockpit has already done the thinking, so a dean gets the headline in ninety seconds without touching a filter.
Drill all the way down. Same entry point, but they pull the funnel apart by cohort, check it against peers, and leave with an action.
Information architecture
One platform, five tabs, one vocabulary.
The fix was structural, not decorative. Learn one page and you can read them all.
Data sources
Slate
undergrad + graduate admissions
Student records system
enrolled-student records
Shared metric definitions
one meaning per number
Meridian Institute Analytics
one platform, one shared vocabulary
Overview cockpit
all four departments at a glance
Undergraduate
- Summary
- Geo
- Funnel, yield, melt
- Applicant segments
Graduate
- Summary
- Geo, source markets
- Trends
- Definitions
Research
- Summary
- Proposals
- Awards
- Expenditure, faculty
Human Resources
- Summary
- Headcount trends
- Turnover trends
- Workforce composition
Global filters (year, term, level, decision plan, cohort) carry across every screen, so a number always means the same thing.
Layouts before colour
Layouts tested before any colour could rescue them
Options side by side, so the structure had to win on its own merits rather than on styling.
Grayscale only, and always more than one option per screen, because a single wireframe invites approval rather than a decision. I put them side by side in front of stakeholders and asked which answered a specific question faster, not which they preferred. Anything that needed me to explain it lost.
Overview, the institutional cockpit
Undergraduate, Summary, with the world map
Undergraduate, Application Totals
Graduate, Summary, with the world map
HR, Summary, plain and protective
HR, Trends Headcount, ten-year combo
Research, Summary, money over time
Research, H-Index distribution
Undergraduate
A tension, not a headline.
- Applications fell, but acceptance went from 43.0% to 51.2% and yield from 17.0% to 21.0%.
- A bigger share of a smaller pool, converting better. The tab is built around that funnel.
- Geography sits at the top level, not in a sub-chart. College-age population is shrinking in the Northeast and growing in the South.
Graduate
Same layout, completely different engine.
- Small pool, heavy international skew, and two thirds of admitted students never enrol.
- So this tab is built around stage conversion rather than application volume.
- India is the largest source country, which means a visa policy change is not a news story here. It is an enrolment event.
- One shared admissions template would have halved the build, and the team wanted it. I pushed back, because merging the two recreates the exact problem I was there to fix.
Research
The highest-value fix was almost embarrassingly simple.
- Three numbers kept getting confused: funding won, awards currently managed, and money actually spent.
- The old dashboard put them in three separate pie charts, so people were comparing slices across charts. Humans are genuinely bad at that.
- I put them side by side in one band with one clear definition each. No new data, no clever visualisation.
What I could not change
The fixed points I designed around.
Two minutes and one hour, on the same screen
A dean skims. An analyst excavates. Two interfaces would have split the vocabulary again, so I built one: the headline answers in a glance, and every tile is a door into the detail underneath it. Nobody gets a lesser version.
Designing to what the platform can actually render
Power BI will not do custom components, so half of what I could draw was unbuildable. I set the palette, spacing and chart rules inside its native visuals first, then designed. Nothing in the file needed a workaround to ship.
A screen that has to survive missing data
Demographic panels go blank when a category cannot be released or a count is too small to publish. Rather than let those views break, empty is a designed state: the panel says what is suppressed and why, and the layout holds.
The platform
One platform, read top to bottom.
Eight screens, five tabs, one shared vocabulary.
What changed
Six rounds of testing, and the numbers held.
of leadership using it inside the first term
departments retiring their old dashboard within a quarter
definition disputes once the shared dictionary is in place
Before figures were measured across the four legacy dashboards during the audit. After figures come from six rounds of validation with real users and stakeholders, timed and counted in those sessions rather than projected.
First walkthrough
None of this survived the first walkthrough intact.
I ran the work through six rounds of walkthroughs with real users and stakeholders, clickable at wireframe stage and again at final design, before anyone signed anything off. Four changes came out of it that I would not have arrived at on my own.
Clickable prototypes both times, so people navigated instead of nodding at a picture. I gave each person a real question from their own job and watched them try to answer it, timing them and staying quiet. Wireframe stage caught the structural problems while they were still cheap; the final-design round caught the language, which is where most of the four changes came from.
The overview was answering four questions, not one
First walkthroughs stalled on the landing screen: people scanned it like a report instead of reading it. I cut it to a single question, how is the institution doing, and pushed everything else one level down.
Nobody trusted a number they could not trace
Analysts kept asking where a figure came from. Every metric got a definition on hover and a visible source, and the objections stopped.
Year-on-year change was being read backwards
A falling number with a green arrow beside it confused almost everyone, because down is good for melt and bad for applications. Direction now follows the metric, not the maths.
Filters kept resetting between tabs
Testers lost their cohort every time they moved across the platform and had to rebuild it. Filters became global and persistent, which is the change people mentioned most in the final round.
What I took from it
The hardest work was not visual at all.
- Define before you draw. Most of this project was agreeing what a number is allowed to mean. The layout was the easy half.
- Win the sceptics first. I designed for the analysts who defend this data to a board. Once they trusted it, everyone downstream did too.
- Constraints belong in the room early. I lost a week to layouts the platform could not build. The engineer now joins at wireframes, not at handover.
- Say what you have not proven. The AI explains, it never scores, and it is not bias-audited yet. Naming that earned more trust than hiding it would have.
Thank you
Thanks for reading.
Happy to walk through any part of this in more detail, including the decisions that did not make it.
