Case Study. Governed data lake, dashboards and an AI analyst
A governed data lake, live dashboards and an AI analyst.
A multi-site healthcare network ran its reporting as dozens of hand-run scripts and a weekly KPI workbook nobody could reproduce. We built a governed data lake and showed it working on their own data first. It has since become their reporting system of record: 28 reports reconciled to their own method, an AI analyst staff ask in plain language, marketing joined to the clinical record, and native iPhone, iPad and Android apps.
- Zero extraction errors
- Reconciled to about 1%
- Your own data first to first demo
- Security and governance built in
The situation
Everyone trusted the workbook. Nobody could reproduce it.
The client is a multi-site healthcare network running two core systems plus three retail locations. Leadership met weekly against a KPI workbook that one person maintained by hand, downstream of dozens of scripts that were run manually before every meeting.
Nothing was wrong with the numbers. The problem was that no one could re-derive them, nobody could ask a question that was not already in the workbook, and the two systems describing the same customers shared no common identifier, so the most valuable question in the business could not be answered at all.
We are not naming the client. Everything below is aggregate and verified.
What we built
Four layers, and two security boundaries.
We read from your systems on a schedule and never write back to them. What comes across is agreed with you up front, and who can see it is decided by role.
How it ran
Three phases from first access to a live demo.
Phase 1. Access
Read-only access over a private connection, agreement on exactly which fields were in scope, and the first refreshes running. Not one extraction error.
Phase 2. Their definitions
The client's own measures rebuilt and reconciled against their workbook week by week, 45 weeks of history, before anyone was shown a dashboard.
Phase 3. Dashboards and analyst
The scorecard live, plus heatmaps, catchment maps and forecasts, and the analyst held to the approved views and the aggregation rules.
The audit
Before the client's own team demonstrated it, every headline number was re-derived independently from the raw data. Ten of ten matched exactly.
Results
Measured, not estimated.
Coverage
Every system of record
Clinical, billing and point of sale brought into one governed lake and reconciled together, refreshed nightly with zero extraction errors. Integrating separate systems is the hard part, not the size of any one of them.
Accuracy
1%
average variance against the client's own weekly KPI workbook across 45 weeks, with the residual explained by late-posted bills.
Independent audit
10 of 10
headline numbers re-derived independently, straight from the raw data. All ten matched exactly.
Forecast error
7 to 18%
one week ahead on the main KPIs, stated on the chart. The 80% band contains the actual value 82% of the time.
Proof before commitment
Your own data first
from first database access to a live demo on the client's own real data.
The view nobody had seen
Where the patients actually live.
Catchment by location, two lines of business on one map, area names rather than codes. In the room, this was the slide that changed the conversation from reporting to planning.
The question nobody could answer
Two systems, no shared key.
The two core systems described the same people and had nothing in common to join them on. We measured how far they matched, took out the matches that were pure coincidence, and reported what was left.
It answers how many, never who. The result: roughly three quarters of one line of business were also customers of the other, a number the organization had never been able to state.
Security and governance
Controls set to your policy, not ours.
Cloud hosted and managed by us, with private connectivity to your source systems and MFA enforced before any page is served
Read-only access to the systems we read from. The platform never writes back to them
Role based access decides who sees what, so your teams get the depth they need internally while sensitive fields stay restricted
The analyst reads only the views you approve, never writes, and shows the query behind every answer, so any number can be traced back
Aggregation and suppression thresholds are configured to your own policy, including none at all where your teams need detail
A saved report is checked against the current rules every time it is opened, so it cannot outlive the policy it was written under
One more thing
Attributes they did not know they had.
The source system carried no standard field for a set of attributes the network cared about. Staff had been recording them in custom fields anyway, for years, for roughly half the customer base.
Nobody had been able to use them because nobody could get at them at scale. They are now on a dashboard.
Reporting
Every report, rebuilt and proven.
The client's hand-run reports were rebuilt one by one as live pages, and each was checked against the client's own numbers before anyone relied on it. 28 report pages now cover clinic visits and providers, who the clinics see and where they live, the diabetes education program (educators against target, pumps and CGM, workshops, the documentation audit), foot care (revenue, lead sources, invoices, productivity), pharmacy patients, follow-up care, a clinical cohort study and email marketing.
Reconciled to the client's own method.
28
live report pages, each rebuilt and checked against the client's own numbers.
about 1%
the diabetes education audit, a complex report built by hand each week, with every step of the difference reconciled.
0.4%
the lost-to-follow-up report against the client's own query.
Any range
every report takes any date range, every chart can be redrawn, and every page downloads as Excel with native, editable charts.
Marketing and care
Marketing joined to the clinical record.
The client's email marketing CRM is exported to the lake nightly and linked to patients, so the platform answers what neither system could alone: who email can reach at each clinic, what happens to bookings in the days around each email, and how email reach crosses from one service line to another.
The analysis is careful about cause and effect. People who open emails were already more engaged before they opened them, so the platform compares like with like rather than crediting every booking to the email.
Research
Clinical research on operational data.
The platform ran a retrospective cohort study on the network's own records: patients who started a continuous glucose monitor, against a matched comparison group. The effect was estimated inside matched bands and then checked several ways: excluding patients who started a new therapy, matching on further medication, looking at twelve months, and splitting by device.
The finding belongs to the client, so it is not published here. What it shows is that a well-governed operational lake can support research-grade questions, not just weekly reporting.
The analyst
Now part of the team's routine.
Staff ask Spella in plain language, by text or by voice, and the answers come from the same governed numbers as every report.
Any answer can be saved as a report, kept private, shared with chosen colleagues, or shared with the whole organization
Only its owner can rename, re-share or delete a saved report
Every list, page, download and the mobile app honor the sharing choice
She says dates in full when she speaks, such as October 8th, a small thing users notice
Mobile
In their pocket: native apps for iPhone, iPad and Android.
One codebase, built in-house and distributed to staff through TestFlight, carrying every one of the 28 web reports with the full content of each page.
What the apps do.
Maps of every postal area with pinch to zoom, heatmaps, flows, cohort charts and tables
Any chart opens full screen and turns sideways for more room
Spella by voice: hold, ask, and hear the answer, then save and share it from the phone
Face ID unlock, and sign-in with the same staff account as the web
Totals only on the device: no patient name, identifier or list is ever sent to a phone, whatever the person's access on the web
Loading states shaped like the content and an updating view when a filter changes, so the app never shows a blank screen
Operations
Run like a product, not a project.
Staff accounts are created from the platform, with a branded invitation sent from the client's own email domain
The sign-in page and every email carry the client's brand
Patient-level access is granted person by person
The report that lists patients by name is gated like every other
See it on your own data
The same four layers work on claims, ERP, CRM, point of sale, subscription platforms and spreadsheets. Tell us which systems hold your numbers and what leadership asks for every week.