Business Challenge
- 30-day readmissions attract NHS penalties worth millions annually
- Premature discharge causes complications; delayed discharge wastes capacity
- Clinical judgement alone cannot process hundreds of simultaneous risk variables
← Showcase
4.1 · Showcase
Interactive Projects: From Concept to Application
Three end-to-end projects spanning healthcare, finance, and retail — each showing the full journey from business problem to deployed solution, with interactive demos that let you adjust parameters and see results in real time.
Each project tells a complete story: a real business challenge, the technical decisions made to address it, and measurable outcomes. Transparency about methodology and trade-offs is deliberate — in production, reliability and interpretability often matter more than squeezing out the last percentage point of accuracy.
Project 1 · Healthcare
Healthcare Readmission Risk Predictor
Identifying high-risk patients before discharge so targeted interventions can prevent costly and harmful readmissions.
Technical Stack
- XGBoost trained on 50,000 anonymised patient episodes (2 years)
- Class imbalance (15–20% readmission rate) handled with SMOTE
- Features: demographics, clinical measurements, temporal patterns, medication complexity, social determinants
- SHAP values provide per-patient explanation for clinicians
Results
- 78% recall at 65% precision — catches ~4 in 5 future readmissions
- 30% readmission reduction among high-risk patients
- £312k value per 1,000 discharges (conservative estimate)
- Piloted successfully at 3 hospital sites
0%
Recall @ 65% Precision
0%
Readmission Reduction (high-risk)
£0k
Value per 1,000 Discharges
0
Hospital Sites
Key Insight — Pilot sites cut readmission rates by 30% among high-risk patients, with the model identifying nearly four out of five who would readmit (78% recall at 65% precision) and creating value exceeding £312,000 per thousand discharges.
Challenges
Missing social determinants
Inconsistently recorded in historical records — resolved with multiple imputation and improved collection protocols.
Clinician resistance
Overcome by implementing SHAP values so clinicians could see exactly which factors drove each individual risk score.
Project 2 · Finance
Financial Fraud Detection with Graph Neural Networks
Detecting fraud rings by analysing the structure of transaction networks — catching patterns that per-transaction methods miss entirely.
Business Challenge
- Billions lost annually to fraud across UK financial institutions
- 30–40% of declined legitimate transactions cause customer switching or abandonment
- Fraudsters use mule accounts, rapid multi-hop money movement, and circular transactions — invisible to per-transaction methods
GNN Approach
- Graph Attention Networks — account connections reveal fraud rings
- Bipartite graph: accounts and transactions both as nodes with directed edges
- Trained on 10M transactions (2 years); 0.2% fraudulent; focal loss for class imbalance
- Temporal validation: train months 1–18, validate 19–21, test 22–24
- Node features: centrality, clustering coefficients, temporal velocity
Deployment
- GraphSAGE sampling (15–25 neighbours/hop) for scalability
- Redis caching of frequently accessed graph regions
- Kubernetes; ONNX Runtime + quantisation
- Sub-100ms inference at scale
- GNNExplainer surfaces suspicious subgraphs for investigators — satisfies regulatory requirements
| Approach | ROC AUC | Precision @ 75% Recall |
|---|---|---|
| Rule-Based System | 0.71 | — |
| Standard ML (gradient boosting) | 0.78 | 65% |
| Graph Neural Network | 0.89 | 85% |
0%
Precision @ 75% Recall
0%
Fraud Loss Reduction
0%
Fewer False Positives
0.89
ROC AUC (vs 0.78 gradient boosting)
Key Insight — Analysing account relationships rather than isolated transactions raised ROC AUC from 0.78 to 0.89, cutting fraud losses by 45% whilst simultaneously reducing false positives by 28% — fewer customer declines alongside better fraud prevention.
Project 3 · Retail
Retail Demand Forecasting with Hierarchical Time Series
Forecasting demand across 10 million product-location combinations whilst keeping every level of the hierarchy mathematically consistent.
Business Challenge
- 50,000 products × 200 stores = 10M product-location forecasts required
- Independent forecasts are incoherent: store totals don't sum to company total
- Incoherence makes supply-chain planning impossible
- Low-volume / intermittent demand defeats standard forecasting methods
Solution
- Facebook Prophet as base model (multiple seasonality, holiday effects, external regressors)
- Regressors: promotional calendars, weather, holiday indicators, competitor activity
- MinT reconciliation combines bottom-up, top-down, and middle-out forecasts
- Low-volume products: ensemble of Prophet + exponential smoothing + seasonal naïve, weighted by historical performance
Results
- 22% reduction in mean absolute error vs independent forecasts
- 15% reduction in safety stock requirements
- 12% improvement in product availability (fewer stockouts)
- £8M annual value for pilot retailer
0%
MAE Reduction
0%
Safety Stock Reduction
0%
Availability Improvement
£0M
Annual Value (pilot retailer)
Lessons & Future Direction
Simplicity wins in production
Complex neural nets achieved marginally better offline accuracy but proved fragile and hard for users to trust. Prophet + hierarchical reconciliation delivered similar results with far better maintainability.
Future enhancements
- Causal impact analysis for promotions and price changes
- Cross-store learning to bootstrap new-product forecasts
- Integration with assortment optimisation
Key Takeaways
- Interactive projects demonstrate the complete journey from business challenge to deployed solution
- Each project spans different industries, techniques, and complexity levels
- Transparency in methodology, challenges, and results builds trust
- Real-world deployment requires balancing performance with interpretability and operational practicality