Sample projects

See the work before you hire us

We're a new company, so instead of client logos we show sample projects: real dashboards and models, built for realistic scenarios using made-up data.

These are sample projects, not client work. The organisations are fictional and every number is synthetic (made up by us) unless a project says otherwise. No real company's data, name or results are used. As we complete real projects, with our clients' permission, we'll replace these with real case studies.

Sample project · Analytics & BI · Synthetic data

Loan portfolio dashboard for a Lagos microfinance bank

The scenario

A fictional microfinance bank with six branches across Lagos. Its loan book is tracked in spreadsheets, and management gets one bank-wide report each month, about two weeks after month end.

The question

Where are repayments slipping, and early enough to act?

What we built

A dashboard that tracks portfolio at risk (PAR30: the share of loans more than 30 days overdue) by month and by branch, against the bank's 5% target.

Gross loan portfolio
₦2.06bn
PAR30, all branches (Oct 26)
4.8%
Branches above the 5% target
1 of 6
PAR30 by month: All branches
PAR30 by branch, October 2026

What it reveals

In this sample data, the bank-wide figure is 4.8%, still under target, so the usual monthly report looks fine. Split by branch, Ikorodu stands out: PAR30 climbs from about 4% in May to 11.2% in October, while every other branch stays under 5%. That pinpoints one branch and one point in time for management to look into, such as a change in loan sizes or staffing, months before the bank-wide number crosses the target. Try it: choose Ikorodu in the branch filter.

Data used: synthetic. Twelve months of branch-level figures made up by Bloom InfoSys for this sample. The branch names are Lagos areas only and don't refer to any real bank.

Sample project · Predictive model · Synthetic data

Customer churn early-warning model for a mobile wallet

The scenario

A fictional fintech whose customers often stop using the wallet without closing their account. Retention offers go to everyone, at a high cost.

The question

Which customers are likely to stop in the next 90 days, so retention offers go where they count?

What we built

A model that gives every customer a risk score from their recent activity, ranks them into ten equal groups (deciles), and explains which signals drive the score.

Customers in the sample
20,000
Left within 90 days
10.9%
Leavers in the riskiest 20%
57%
Share who left, by risk decile (1 = riskiest)
What drives the score (share of predictive power)

What it reveals

In this sample data, the 20% of customers the model ranks riskiest include 57% of everyone who left. A retention team that focuses there reaches most likely leavers while contacting a fifth of the customer base. The strongest signals are behaviour changes, such as days since the last transaction and falling top-ups, which a team can watch for every week.

Data used: synthetic. 20,000 made-up customer records with activity patterns designed to resemble a mobile wallet. On a real project, the model is trained and tested on your own history, and we report its accuracy on data it hasn't seen.

Sample project · Healthcare analytics · Synthetic data

Claims review analytics for an HMO

The scenario

A fictional HMO that processes claims from 40 hospitals and clinics. Claims are checked by hand, and the team can review only a small share each month.

The question

Which claims, and which providers, should reviewers look at first?

What we built

Automated review rules that flag claims for a closer look, plus a provider view that shows where flagged amounts concentrate. Flags are for review, not accusations.

Claims value, 12 months
₦1.8bn
Flagged for review
₦116m (6.4%)
Flagged amount from 3 providers
47%
Amount flagged for review, by rule (₦ millions)

What it reveals

In this sample data, 6.4% of the claims value is flagged, and the biggest single rule is cost far above what similar providers charge for the same service. Three of the 40 providers account for 47% of the flagged amount. Instead of sampling claims at random, reviewers start with those three providers and the highest-value flags, then confirm each case with the provider before any decision.

Data used: synthetic. Made-up claims, providers and plan rules. No patient or provider information of any kind was used.

Want this for your own data?

Start with a free first call. If a Data Health Check is right for you, you'll have a fixed quote within 3 working days.

Book a first call