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

Classification, clustering and representation learning on audio and tabular data, from feature design to honest evaluation. Currently completing the University of Cambridge Data Science Career Accelerator.

Open to machine learning and data science roles, and to PhD positions in machine learning for audio.

See the work Download CV
Fig. 1 · Student dropout. Dropouts found by the tuned XGBoost model as student data arrives in three stages. Held-out test set: 5,013 students, 779 dropouts.
Selected work

Five projects, each with the number it reached

Students still on course who later dropped out, Stage 3 model.
Supervised learning · University of Cambridge · 2026

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 showed the decision threshold would need recalibrating for each new intake.

Python · XGBoost · Keras · scikit-learn · pandas

0.977 ROC-AUC on students still on course
Accents 3 · 6 · 8 · 10 · 12 · felt downbeat: 3 · Mode: Phrygian
Music information retrieval · 2026

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. Predictions under 0.4 confidence are flagged as uncertain.

librosa · scikit-learn · Streamlit · pytest

Live demoCode

0.645 F1-macro, cross-validated (±0.115)
ROC over 2,418 test–train recording pairs, 78 of them same-singer.
Deep learning · audio · 2026

Recognising a singer's voice rather than their style

109 historical recordings of 79 cantaores, only 31 of them with more than one recording: too little data to train an encoder from scratch. I fine-tuned a projection head on pretrained OpenL3 embeddings with triplet loss 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

0.80 AUC, closed-set verification (EER ≈ 30%)
Customers per segment, K-means with k = 5.
Unsupervised learning · University of Cambridge · 2026

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. The elbow method, silhouette score and Davies–Bouldin index all pointed to k = 5, and hierarchical clustering confirmed the split.

Each segment comes with a marketing action, from retaining the champions to winning back customers who had not bought in three years.

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

Code

68,300 customers in five segments
In progress
Natural language processing · University of Cambridge · 2026

Classifying sentiment in text

Current project in the Cambridge NLP module. Method and results will be added here when it is finished.

Python · NLP

In progress
Method

How I work, in four steps

The same sequence runs through every project above. Each step has an example from the work.

  1. 1 · Data

    Audit before modelling

    Check what the data can and cannot say. In the dropout data, 182 students appeared twice, so the split was grouped by student.

  2. 2 · Features

    Encode what the domain knows

    Mode and compás position for flamenco. For dropout, the order in which information about a student actually arrives.

  3. 3 · Model

    Start from a baseline

    Simple models first. In the dropout project, tuning moved ROC-AUC by less than 0.003. The data mattered more than the configuration.

  4. 4 · Evaluation

    Test the result against itself

    Ablations, probes and temporal splits. Removing nationality cost 0.030 AUC, and a temporal split cut early recall from 0.54 to 0.35.

Listen

Play a few notes. This is what an audio model sees.

Each key you press is drawn as a spectrogram: time runs left to right, pitch goes up, brightness is energy. The palo classifier and the singer embeddings both start from time–frequency pictures like this one.

3 kHz 100 Hz Andalusian cadence · A G F E

Click the keys or type A S D F G H J K. Synthesised in the browser with the Web Audio API, no samples.

Research interests

Machine learning for audio and music

Representation learning, music information retrieval, and how what a model separates relates to what a listener actually perceives. My training as a classical and flamenco pianist is where the features and the questions come from.

Education
  • 2026Data Science Career Accelerator, Machine Learning & AIUniversity of Cambridge (PACE)in progress
  • 2025Master's in Neuromusic: Music–Brain InteractionUniversidad Católica San Antonio de Murcia (UCAM)60 ECTS
  • 2023 – 2024Master's in Advanced Studies in Flamenco, Flamenco PianoEscola Superior de Música de Catalunya (ESMUC)9.5 / 10
  • 2022Título Superior de Música, Classical Piano (Bachelor's equivalent)Escola Superior de Música de Catalunya (ESMUC)9 / 10
Contact

Roles in machine learning and data science, PhD positions, collaborations.

laura.ortega.garre@gmail.com

LinkedIn GitHub CV (PDF)