I build machine-learning models, and report what they actually get right.

Predicting student dropout, and showing when the prediction is useful

25,059 international students, with data arriving in three stages: applicant details, then attendance, then module results. I compared XGBoost with a neural network at each stage, using a split grouped by student and a held-out test set of 5,013.

The late-stage score of 0.993 was too good: imputed academic records were giving the answer away. Re-evaluated on the 4,339 students still on course, the model finds 105 of 136 future dropouts. A temporal split kept ROC-AUC at 0.891, 0.904 and 0.980 but showed the decision threshold would need recalibrating for each intake.

Python · XGBoost · Keras · scikit-learn · pandas

Separating network-wide complaints from site-specific ones in customer reviews

Tens of thousands of reviews from two platforms for a UK fitness chain, a client under NDA. I compared BERTopic (MiniLM embeddings, UMAP, HDBSCAN, hierarchical merging), transformer-based emotion classification, LDA tuned on c_v coherence, and a local LLM (phi4-mini on Ollama, temperature 0) for theme labelling.

Results are reported as rates with minimum-volume thresholds instead of raw counts. I corrected date artefacts from bulk uploads and documented where the LLM's labels were vague or unsupported by the data.

BERTopic · sentence-transformers · UMAP · HDBSCAN · Hugging Face · Ollama · Gensim

Turning 951,668 orders into five customer segments

Built a customer-level view from raw order data: frequency, recency, lifetime value, unit cost and age for 68,300 customers in 47 countries, in a scikit-learn pipeline with StandardScaler and K-means.

I chose k = 5: the elbow flattens there and the silhouette score peaks, while the Davies–Bouldin index was marginally better at k = 6, which the analysis reports. Ward clustering showed a broader two-group structure that the five segments subdivide, and each segment comes with a marketing action.

pandas · scikit-learn · SciPy · PCA · t-SNE

Code

Does musical surprise explain how the brain responds to Bach?

Public CC0 dataset: EEG from 20 listeners, 7,839 notes. I trained 24 GRU sequence models in PyTorch (pitch and rhythm, three seeds, with and without augmentation), with the design fixed before training, and fitted temporal response functions in MNE with cross-validation nested by piece.

Pitch surprisal from the augmented GRU improved EEG prediction in 16 of 20 listeners, more than IDyOM or IDyOMpy, and held across seeds and when leaving out a piece or a listener. The rhythm result did not hold, and the write-up says so. Along the way I reproduced a published effect and showed that trial-wise validation inflates it.

PyTorch · MNE-Python · ridge regression · Wilcoxon tests · pytest

Code on request

Telling flamenco palos apart from audio

Classifies soleá, seguiriya, bulería and alegrías. The four share a 12-beat compás, so the features come from music theory: Phrygian versus major mode, the Andalusian cadence, and where harmony falls within the cycle, alongside MFCCs and chroma.

A Random Forest beat gradient boosting and an SVM. Tempo turned out to be the strongest single feature, and the beat-position features lowered cross-validation variance rather than raising the mean. Deployed as a Streamlit app, with pytest tests and GitHub Actions CI.

librosa · scikit-learn · Streamlit · pytest

Live demoCode

Recognising a singer's voice rather than their style

109 historical flamenco recordings of 79 singers, only 31 of them with more than one recording: too little data to train an encoder from scratch. I trained a 512-256-128 triplet-loss projection head on pretrained OpenL3 embeddings and evaluated closed-set verification.

The fine-tuned space beats the frozen baseline. A follow-up probe showed it also separates the two source datasets, so part of what it learns is recording origin, and the write-up says so.

PyTorch · OpenL3 · triplet loss · UMAP

Code

Tracking airline disruptions the way an operations centre would

Flask REST API with Pydantic validation and API-key authentication, SQLAlchemy and PostgreSQL with Alembic migrations, and a Celery worker that estimates impact and propagates delays along an aircraft's rotation. Runs in Docker Compose, with a GitHub Actions pipeline that tests the full stack end to end.

Flask · SQLAlchemy · PostgreSQL · Celery · Redis · Docker · pytest

Code