Featured Work

Selected projects

Client work and Cambridge coursework. Each case study states the method, the measured result and its limitations.

Featured

Financial Regulation / NLP
G-SIB Disclosure Analysis Prototype
81 reports and transcripts 17,292 text chunks scored
Natural Language Processing
Customer Review Topic Modelling and Emotion Analysis
27,585 reviews analysed 2,031 anger-led reviews
Anomaly Detection / Operations
Ship Engine Anomaly Detection
422 readings flagged (2.16%) IQR rule
Applied AI / RAG / Privacy
Private Local Documents AI
25+ file formats no third-party AI provider

Additional Projects

Further work spanning forecasting, deep learning, predictive analytics, and QA engineering.

Forecasting
Book Sales Demand Forecasting
Weekly demand for two long-selling titles forecast with SARIMA, XGBoost, LSTM and hybrids on a 32-week hold-out. A well-specified seasonal model was the strongest default; the hybrids did not beat it.
19.6% MAPE, SARIMA (Title B)
Predictive Analytics
Student Non-Completion Prediction
XGBoost at three stages of the student journey. Adding engagement data gave the best balance of accuracy and lead time; the final-stage result is flagged for a leakage check.
0.650 F1, engagement stage
Time Series
Crime Incident Forecasting with SARIMA
Daily incident data for a US city resampled, tested for stationarity and modelled with seasonal ARIMA, with a 52-week cycle confirmed and models compared on a hold-out year.
139.98 test MSE, best SARIMA
Financial Services
Bank Customer Churn Classification
Decision tree and random forest models predicting which retail bank customers will leave, with the trade-off between catching churners and raising false alarms made explicit.
0.61 churner F1, random forest
Statistical Analysis
Statistical Hypothesis Testing Framework
Six business questions tested with t-tests, ANOVA, chi-square and Mann-Whitney U, with normality checks driving test choice and three nulls honestly retained.
6 hypothesis tests at α = 0.05
Deep Learning
Neural Network Design and Hyperparameter Search
A Keras spam classifier built, tuned and evaluated over three exercises. Grid search over epochs and batch size did not beat the simple 64-32 baseline.
94.6% baseline test accuracy
QA Engineering / Test Automation
Automated PDF Compliance Test Harness
Layout and content checks for generated PDF letters, with per-file failure reasons and a colour-coded Excel report. Verdicts matched ground truth on a synthetic 300-letter corpus with five seeded defect types.
300 / 300 verdicts matched (synthetic corpus)
Business Analytics / ML
Customer Segmentation with Clustering
K-Means and hierarchical clustering of 68,300 retail customers, published as a learning exercise: silhouette scores were modest and one cluster held two-thirds of customers.
0.266 best silhouette score
Statistical Analysis / Insurance
Medical Insurance Cost Correlation Analysis
Pearson correlation analysis of what drives insurance charges across 1,000 anonymised records, set in a scenario where an employer weighs a BMI-linked benefits scheme.
r = 0.709 BMI correlation, strongest predictor identified
Marketing Analytics / ML
Decision Trees and SHAP: Bank Marketing
Tree-based models and SHAP on a public bank telemarketing dataset. Call duration dominated, but it is only known after the call, so the score describes outcomes rather than pre-call targeting.
0.9485 ROC AUC, Gradient Boosting (with leakage caveat)

Additional Work

Broader technical exploration outside my core financial services focus - methodology deep-dives and applied ML across other domains.

Natural Language Processing
Movie Review Sentiment Analysis
A classical baseline on binary sentiment, where Bag-of-Words and TF-IDF tied, then four recurrent networks on a harder five-class task, where the LSTM and GRU failed to learn.
38.5% five-class accuracy, Bidirectional LSTM
Business Analytics / Retail
Customer Loyalty Analysis
Regression and correlation analysis across 1,711 customers isolating which of quality, brand awareness and negative publicity actually drives loyalty.
r = 0.713 quality-loyalty correlation
Feature Engineering / Automotive
Automobile Price Prediction
Feature engineering, feature selection and PCA/t-SNE on a 205-vehicle dataset, comparing Random Forest and linear regression on a held-out test set.
R² 0.894 Random Forest, held-out test
Deep Learning Fundamentals
Neural Network Manual Backpropagation
One forward and backward pass of a small network written by hand in NumPy, without a framework, to show exactly what TensorFlow or PyTorch automates.
Zero framework dependencies
Deep Learning / Model Evaluation
Optimiser Comparison: Adam vs RMSprop
Adam and RMSprop compared on the same small network for historical passenger-survival data, then regularisation, dropout and early stopping explored.
86.0% vs 83.2% test accuracy, RMSprop vs Adam