Curiosity first.
Clarity follows.

1. Featured study01 / 08

   01Clustering2025

   Hotel customer segmentation

   Which hotel guests behave differently, and how could the hotel tell them apart?

   records cleaned and analysed

   83,590

   clustering methods compared

   3

   ways of plotting the groups

   2

   • Python
   • pandas
   • scikit-learn
   • K-Means
   • DBSCAN
   • HDBSCAN
   • PCA
   • t-SNE

   Code for Hotel customer segmentation on GitHub (opens in a new tab)

   Inside the projectProblem, method, sources

   Problem

   The hotel had 83,590 customer records, and a lot of them needed work before they were useful. Some columns were hashed IDs, revenue was split across two channels, and a field called DaysSinceCreation was really measuring how long someone had been a customer.

   Method

   1. I dropped the ID, NameHash and DocIDHash columns. They say nothing about how a guest behaves, and keeping them would have grouped people by identity instead of behaviour.
   2. I explored revenue, lead time, cancellations, no-shows and special requests before deciding which features to use.
   3. I clustered the data three ways. K-Means gave me compact groups, and DBSCAN and HDBSCAN picked up the unusual stay lengths and big spenders that averages hide.
   4. I used PCA and t-SNE to plot the groups in two dimensions, so I could actually see them and sanity-check them.
   5. I wrote up each group as a customer profile, with a suggestion for how the hotel could look after it.

   Sources

   • 83,590: HotelCustomersDataset.xlsx
2. SQL · Database design02 / 08

   02SQL · Database design2025

   Booking and retail databases in SQL

   Two MySQL databases that keep their own data clean, plus the queries I used to analyse them.

   databases designed from scratch

   2

   triggers and stored procedures

   5+

   • MySQL
   • Triggers
   • Stored procedures
   • CTEs
   • Window functions

   Code for Booking and retail databases in SQL on GitHub (opens in a new tab)

   Inside the projectProblem, method

   Problem

   If a booking database has mistakes in it, every report built on top of it has the same mistakes. Ages go out of date, a changed passenger ID leaves other tables pointing at nothing, and one duplicated employee can double a headcount.

   Method

   1. For the train booking system I built three linked tables, with triggers that work out each passenger's age from their date of birth and update IDs everywhere when they change.
   2. I added a five-character rule for Passenger_id and proper foreign keys, so bad rows can't get in at all.
   3. I made backup tables and an EMPTY_DATA() procedure, so I could reset a test run with one call.
   4. For the retail store data I ranked stores and employees, averaged sales by product, and matched up staff records using joins and DISTINCT.
   5. I removed duplicate employees directly in the table instead of copying the clean rows into a second one.
3. Business intelligence03 / 08

   03Business intelligence2025

   Customer churn dashboard

   A Power BI report that shows not just how many customers left, but who they were and where they were.

   views in one report

   6

   breakdowns: demographics, plan, geography

   3

   • Power BI
   • Power BI Service
   • DAX

   Code for Customer churn dashboard on GitHub (opens in a new tab)

   Inside the projectProblem, method

   Problem

   A single churn number doesn't tell a manager much. They need to know which customers left, where they were and what they had in common. Without that, the report gets a nod in a meeting and is never opened again.

   Method

   1. I used a public churn dataset from Kaggle, so anyone can check my numbers against the source.
   2. I showed churn over time rather than as one figure, so there's always something to compare against.
   3. I broke the rate down by demographics, subscription plan and geography to see where it was highest.
   4. I kept every chart cross-filterable and exportable, so people can dig into it themselves.
4. Deep learning · Classification04 / 08

   04Deep learning · Classification2024

   Telecom churn prediction

   Most customers don't leave, so a model can look accurate while missing the ones who do. I built this one to catch them.

   used to balance the training data

   SMOTE

   metrics used to judge the model

   5

   • Python
   • PyTorch
   • scikit-learn
   • SMOTE

   Code for Telecom churn prediction on GitHub (opens in a new tab)

   Inside the projectProblem, method

   Problem

   Telecom churn data is very unbalanced. A model can get a high accuracy score just by predicting that nobody leaves, and learn nothing useful about the customers who actually do.

   Method

   1. I explored the data first, using count plots and density plots to see how each column was spread.
   2. I encoded the categorical columns and scaled the numeric ones, so no single feature dominated training.
   3. I used SMOTE to create more examples of customers who churned, so the model had enough of them to learn from.
   4. I built a neural network in PyTorch and judged it on accuracy, precision, recall, F1 and AUC together, not accuracy alone.
5. Regression05 / 08

   05Regression2024

   Toronto-area house price model

   How well can location, timing and listing activity predict what a house in the Toronto area sells for?

   R² with Gradient Boosting

   0.990

   R² with a neural network

   0.984

   cross-validation

   10-fold

   • Python
   • scikit-learn
   • Gradient Boosting
   • Neural network
   • KNN imputation

   Code for Toronto-area house price model on GitHub (opens in a new tab)

   Inside the projectProblem, method, sources

   Problem

   Toronto-area housing data mixes location, market timing and listing activity in one table. A careless model just memorises the neighbourhoods it has already seen instead of learning how prices actually move.

   Method

   1. I built features around four things: location, sales and prices, listing activity (new listings and the sale-to-list ratio), and timing (year and quarter).
   2. I filled in missing numbers with KNN imputation instead of throwing rows away, and one-hot encoded the categories.
   3. I compared linear regression and decision trees with a Gradient Boosting model, then tried a neural network with early stopping to see if it did any better.
   4. I checked the final model with ten-fold cross-validation rather than trusting a single train and test split.

   Sources

   • 0.990: 0.9898 on the test set
   • 0.984: 0.9844 with early stopping
6. Classification · Ensemble06 / 08

   06Classification · Ensemble2025

   Explainable spam detection

   A spam filter that tells you why it flagged an email, so people can actually trust it.

   accuracy on test data

   96-97%

   models voting together

   3

   engineered features

   15+

   • Python
   • scikit-learn
   • Naive Bayes
   • Logistic Regression
   • SVM
   • Streamlit

   Code for Explainable spam detection on GitHub (opens in a new tab)

   Inside the projectProblem, method, sources

   Problem

   When a real email gets marked as spam with no explanation, people tend to loosen the filter. That's exactly when the dangerous phishing emails start slipping through.

   Method

   1. I combined three models (Naive Bayes, Logistic Regression and SVM), so no single model's blind spots decide the result.
   2. I engineered more than fifteen features beyond the text itself, like word and character counts, links and domains, urgent or money-related words, HTML structure and punctuation.
   3. I added a confidence score, so borderline emails can go to a person instead of being decided by a hard cutoff.
   4. Every prediction comes with a reason in plain English, and the whole thing runs in a Streamlit app that keeps track of its own performance.

   Sources

   • 96-97%: README: “~96-97% on test data”
7. Computer vision · CNN07 / 08

   07Computer vision · CNN2025

   Alzheimer's stage classifier

   It reads an MRI scan and places it in one of four stages of Alzheimer's, instead of just saying yes or no.

   validation accuracy

   99.21%

   test accuracy

   99%

   stages it can tell apart

   4

   • Python
   • TensorFlow
   • CNN
   • Streamlit

   Code for Alzheimer's stage classifier on GitHub (opens in a new tab)Live demo of Alzheimer's stage classifier (opens in a new tab)

   Inside the projectProblem, method, sources

   Problem

   Most image classifiers only say whether something is there or not. With dementia that isn't very helpful, because catching it early matters most, and a plain yes or no can't show what stage someone is at.

   Method

   1. I trained the model to tell four stages apart: non-demented, very mild, mild and moderate.
   2. I built a convolutional neural network and trained it on the Well-Documented Alzheimer's Dataset from Kaggle.
   3. I published the model under an MIT licence and put it online, so anyone can try it on a real scan.

   Sources

   • 99.21%: README
   • 99%: README
8. Reinforcement learning08 / 08

   08Reinforcement learning2026

   Multi-armed bandits, by hand

   I wrote eight bandit algorithms from scratch and tested how they cope when the rewards keep changing.

   algorithms written from scratch

   8

   test environments, 2 of them drifting

   4

   • Python
   • NumPy
   • pytest
   • ruff
   • uv

   Code for Multi-armed bandits, by hand on GitHub (opens in a new tab)

   Inside the projectProblem, method

   Problem

   Bandit algorithms are usually tested in settings that never change, where careful strategies look clever. I wanted to see what happens when the rewards drift over time, because that's what real systems have to deal with.

   Method

   1. I implemented eight strategies myself: epsilon-greedy, UCB, KL-UCB, variance-aware UCB, Beta-Bernoulli and normal Thompson sampling, forgetting Thompson sampling, and sliding-window UCB.
   2. I built test environments where the rewards drift at two different speeds, plus steady ones to compare against.
   3. I wrote one shared evaluation setup with paired statistics, so every algorithm is tested under exactly the same conditions.
   4. I made the whole study reproducible, with pytest and ruff checking the code on every change.

More on GitHub (opens in a new tab)