Financial Services
Mid-market lender · 2024
Machine LearningAnalytics Engineering

Rebuilding credit risk decisioning from the ground up.

Challenge

A regional lender was running a decade-old scorecard that hadn't been recalibrated since the rate environment turned. Default rates were creeping up, the manual override process had become a compliance liability, and nobody could explain why a given applicant was declined.

Approach

We audited the existing model, rebuilt the feature-engineering pipeline, and deployed an ensemble trained on five years of repayment data. Crucially, we framed the objective around approved volume at a fixed loss tolerance — not raw accuracy — and built a profit curve the credit committee could read.

Outcome

Defaults fell 18% while approval volumes held steady. Automated retraining triggers keep the model current, and a regulator-ready model card now satisfies both internal audit and external supervision on the first pass.

Default rate Legacy scorecard vs. new model 100 82 Legacy New model Reduction in defaults −18% ↓ losses
18%
Reduction in defaults
Faster model refresh
100%
Audit pass rate
Retail & Commerce
Specialty retailer · 2024
Machine LearningVisualisation

Demand forecasting across 800+ SKUs in a volatile supply chain.

Challenge

Post-pandemic demand patterns had made a spreadsheet-based forecasting process unmanageable. The merchandising team spent more time on manual adjustments than on decisions, and overstock write-offs were eating into margin.

Approach

We designed a hierarchical forecasting system — bottom-up at SKU level, reconciled top-down to category targets — with external signals (weather, holidays, promotional lift) built in. The whole pipeline runs automatically and feeds the ERP replenishment system directly.

Outcome

Overstock write-offs dropped 22%, manual planning time halved, and the freed-up margin funded the next phase of work. Forecasts now carry honest confidence intervals, so buyers know which numbers to trust.

Overstock write-offs Indexed, year on year 100 78 Before After Overstock reduction −22% ↓ write-offs
22%
Overstock reduction
50%
Less manual planning
4.1%
Gross-margin lift
Healthcare
Private clinic group · 2023
Machine Learning

Reducing no-shows through appointment-level propensity scoring.

Challenge

A 14-location clinic network was running a 19% appointment no-show rate — a direct hit to revenue and a scheduling headache that rippled across every site. Blanket reminders weren’t moving the number.

Approach

We built a real-time propensity model that scores each appointment at booking and again at the 48-hour reminder. High-risk appointments trigger a differentiated outreach workflow. The model sits inside the existing patient-management system, with no change to clinical staff routines.

Outcome

No-shows fell by nearly a third, recovering an estimated £1.4M in annual revenue — without adding a single step to clinicians' day. Schedulers now see risk at a glance and intervene only where it counts.

No-show rate Network-wide, % 19 13 Before After Fewer no-shows −31% 19% → 13%
31%
Fewer no-shows
£1.4M
Revenue recovered
0
Workflow changes
Media & Publishing
Digital publisher · 2023
Machine LearningData Strategy

A subscriber lifetime-value model to guide acquisition spend.

Challenge

A digital publisher was spending aggressively on subscriber acquisition with no clear view of which channels delivered high-value readers versus short-tenure churn. Every channel looked fine on day-one conversion; the truth only showed up months later.

Approach

We built an LTV model that separates predicted engagement from tenure, segments acquired subscribers into behavioural cohorts, and surfaces payback period by channel in a live dashboard for the marketing team. Cohort retention became the lens for every spend decision.

Outcome

The LTV:CAC ratio improved 34%, and the marketing team reallocated budget within 60 days of launch. The strongest cohorts now retain 11 points above the blended average — and the team can see it the moment it happens.

Cohort retentionMonthly subscriber cohorts · % retained by periodM0M1M2M3M4M5M6Jan100827164585451Feb100807063575349Mar100857669646057Apr100796861555047May100888074706663Jun100837366615754LowHighBest cohort (May)M6 retention · best vs. blended+11pp↑ retained
34%
Improvement in LTV:CAC
60 days
To first reallocation
+11pp
Best-cohort retention
Logistics
Last-mile operator · 2022
Analytics EngineeringVisualisation

Consolidating eight data sources into a single operational picture.

Challenge

The operations team worked across eight disconnected systems — route planning, driver apps, customer portals, fleet telematics, and more — none of which agreed with each other. Dispatchers were reconciling spreadsheets instead of running the network.

Approach

We designed a unified data layer, built the transformation models in dbt, and delivered a single operational dashboard giving dispatchers a real-time view across every asset and active delivery. The warehouse now serves as the backbone for all downstream analytics.

Outcome

Eight systems became one source of truth, SLA performance improved 11%, and the operations team finally spends its time on decisions rather than data wrangling. Every later analytics project now starts from this foundation.

Operational systems Sources dispatchers reconcile 8 1 Before After Systems consolidated 8 → 1 one source
8→1
Systems consolidated
11%
SLA improvement
Real-time
Operational view

Have a problem worth solving?

We're selective about what we take on. If your problem is the right fit, so are we.

Get in touch