Analytics product
One view of Europe's audit performance
Financial data sat in separate country-level files with no way
to compare across the continent. I led the build of a single
analytics product that pulls it all together, with automated
ingestion, transformation and validation.
Impact: the first continent-level operational insights
across Deloitte's Audit & Assurance firms, and 4 days saved
every reporting cycle.
PythonSQLAlteryxSynapsePower BI
Read the case study →
Machine learning
Fraud and risk detection model
Suspicious cases were being flagged inconsistently by hand.
Working with business stakeholders, I built a classification
model that scores cases so reviewers can focus on the ones most
likely to matter.
Result: 92% accuracy and 89% recall on evaluation data,
giving reviewers a more consistent starting point.
PythonDataikuscikit-learn
PhD research
Can we predict how well water gets cleaned?
My PhD applied machine learning to predict contaminant removal
in full-scale mine water treatment systems, alongside field
studies of how the seasons combined with design of lagoons and
wetlands affect iron removal.
Published: first author in
Ecological Engineering
(2024) and
Environmental Processes
(2025).
PythonMachine learningField data
Benin Electricity Distribution Co.
Getting the lights back on
There was no customer enumeration team in the branch, so I built
one: trained 20 people, ran the budget, classified customers,
assigned tariff bands and reported to government stakeholders.
Impact: 4,500 customer records collected and power
restored to 1,500 households after 7 years of blackout.
ExcelField data collectionStakeholder reporting
Read the case study →
Automation
From static spreadsheets to live capacity planning
Resource planning ran on static Excel files that went out of
date the moment they were saved. I turned them into interactive,
automated capacity tools.
Impact: clearer view of workloads, fewer scheduling
clashes and better use of specialist people.
AlteryxPower BIExcel
Personal project · Blog
House prices: how many features is too many?
I built and compared five regression models on the Ames Housing
data to find what really drives a home's sale price, and what
happens when you throw every feature at the model.
Finding: Ridge regression won with an R² of 0.78 on
test data. More features only helped once regularisation kept
the model in check.
Pythonscikit-learnRidge regression
Read on the blog →