- Represent the unit faithfully
- Make patterns and differences legible
- Let people export and investigate
03 Selected work / Insights
Enterprise healthcare · Clinical analytics
Clinical & Operational Analytics
Insights.
Making data legible
without prescribing
the decision.
The analytics layer of an enterprise virtual-care platform, spanning ICU and Labor & Delivery. Its most important design decision was where to stop.
Analytics · signal, not noiseTactical now.
Strategic over time.
Strategic over time.
01
The governing position
A mirror,
not an engine.
Insights was deliberately non-actionable: a faithful visual representation of the data, with export to CSV, and nothing more. No recommendations. No “you should.” It showed the unit; it did not tell anyone what to do about it.
The product boundary
- No recommendations
- No implied course of action
- No unearned clinical authority
A mirror reflects. An engine decides. Choosing where a system sits on that line is itself the design work.
02
Two dashboard paradigms
Two clocks.
One analytics layer.
Insights turned what happens inside ICU and Labor & Delivery units into something administrators and leadership teams could read. But those audiences reason on different time horizons.
TacticalNow
The unit,
right now.
Short-term snapshots of live operational data for unit and hospital administrators, with drill-down into a single care unit’s activity.
Shift startLive
08:0010:0012:00Now
StrategicOver time
The system,
over time.
Long-term scorecards, trends, and named comparison across facilities and geographies for leadership making portfolio-level calls.
BaselineDirection
A charge administrator asks, “What is happening this shift?”
A regional leader asks, “Is this hospital improving relative to that one, over the year?”
03
NDA-safe interaction evidence
Same data layer.
Different questions.
Use the horizon switch to move between the unit’s immediate operational view and a portfolio-level scorecard. The reconstruction demonstrates the information architecture, not a production screen.
Single unit · last 30 daysHypertension compliance
FacilityUnitDate rangeApply
Order set compliance
78%Current reporting period
Reassessment timing
42mMedian elapsed time
Severe episodes
16Cases represented
Compliance by monthRepresentative data
NovDecJanFebMar
Concept reconstructionA tactical compliance dashboard for a single condition. Representative data; not a production screen.
For Labor & Delivery, tactical measures included deliveries, vaginal versus cesarean rates, thirty-day readmissions, and the incidence of conditions the clinical protocols were built to manage.
Strategic views pulled those signals into comparison across facilities and geographies. Designing one surface for both horizons would have served neither.
04
The design discipline
Show,
don’t prescribe.
Holding that line took discipline because every dashboard invites the next question: “So what should I do?” The easy answer is to start answering it.
The harder, more honest position was to make the data legible enough that a clinician or administrator could reach their own conclusion—and stop there.
Recommendation in a clinical context carries weight. It implies the system has reasoned about a patient or unit and is willing to stand behind a course of action.
A mirror makes no such claim.Knowing which of those two things you are building, and being honest about it with the people relying on it, is a design responsibility—not merely a product decision.
05
The evolution
A migration is a question,
not just a pipe.
Insights is moving from Tableau toward Databricks, attached to new use cases in perinatal care and ECG acquisition. The current approach is migration-first: move the same information in the same form, gather feedback, then commit to fresh design.
Existing dashboards and predefined workflows
A migration is the cheapest moment to validate the questions, not only replace the pipes.
New platform, perinatal and ECG use cases
This is where I pushed. Re-platforming the same metrics assumes those metrics were right, and I am not convinced they were ever validated against what users actually need to decide.
Analytics platforms impose predefined workflows and layouts, limiting how far a custom interface can go. At enterprise scale, buy-in to move beyond conceptual AI-insights pitches comes slowly. Knowing where the platform ends and design begins is part of the job.
06
What this connects to
Restraint is
a design position.
Choosing to be a mirror rather than an engine is a defensible call with real consequences, not a lack of ambition. The hardest discipline in analytics is resisting the urge to prescribe before you have earned it.
It sharpened the question running through my AI work: when a system moves from showing to recommending, what must it do to earn that authority? Insights chose not to make that leap. Deciding when a system is allowed to is the same problem I now work on with agents.
Next case study
← Previous · Mural