Xoel García Maestu
ES|EN

A Coruña, Spain

Data Science and Engineering

Machine Learning, Deep Learning and Generative Artificial Intelligence.

Data Scientist and Engineer, graduate of the University of A Coruña. I work across the full data lifecycle: capture, preparation, modeling, deployment and visualization.

View projects Résumé
Portrait of Xoel García Maestu

About

I combine data engineering, statistical analysis and artificial intelligence to turn data into useful decisions. From the pipeline that holds a system together to the model that makes it scale, I care about the whole problem and its real impact.

INDITEX / Data & AI Engineer

May 2026 - present / Arteixo, A Coruña · On-site

Data science, data engineering and artificial intelligence within INDITEX's technology team.

NTT Data / Data Analytics & BI Intern

Sep 2025 - Nov 2025 / A Coruña

Technology consulting for banking: Python ETL pipelines (cleaning, integration and validation), synthetic and historical datasets, Looker Studio dashboards and migration of Google Apps Script logic to Python.

Data Engineering

  • Apache Spark
  • Spark MLlib
  • ETL
  • Snowflake
  • PostGIS
  • Denodo
  • Docker
  • AWS
  • Azure

Science and Analytics

  • Polars
  • PySpark
  • ARIMA
  • Power BI
  • Looker Studio
  • Tableau
  • OLAP / MDX

AI, Deep Learning and Agents

  • PyTorch
  • TensorFlow
  • XGBoost
  • U-Net
  • Mask R-CNN
  • VAE
  • HuggingFace
  • MLflow
  • Vertex AI
  • Google A2A
  • Multi-agent systems
  • NLP

Projects

Clinical dashboard: programme activity, organs and recipients, and predicted 5-year risk distribution Clinical platform: assisted donor-recipient allocation, with expected-survival ranking from several models Individual prediction explanation: SHAP/LIME contributions and the tree's decision path Structured registration of a new donor organ, with field validation and example cases

07 / 2026

Bachelor's Thesis: Post-liver-transplant survival prediction

Bachelor's Thesis, awarded a Distinction (Matrícula de Honor, 10/10). It predicts five-year post-transplant survival by combining donor and patient data, on real data from CHUAC (258 cases). The pipeline is complete and reproducible: it imputes missing values (MissForest), selects features and tunes hyperparameters with Bayesian optimization. It trains and compares six models (decision tree, random forest, XGBoost and multilayer perceptron) and adopts the decision tree as the reference for its transparency, validated by a clinician (AUC-ROC 0.757). On top of that, a dockerised clinical web platform (React, Django REST and PostgreSQL) covers the entire transplant lifecycle: registering donors, recipients and transplants, assisted donor-recipient allocation, and an explanation of every prediction (SHAP, LIME, PDP, decision paths and Forest-Guided Clustering). It also adds natural-language search via a local LLM (Qwen2.5 · Ollama). IRLab research group (UDC). The data shown in the screenshots is mocked for the demo. The most sensitive views of the clinical application are not shown.

  • Scikit-learn
  • XGBoost
  • SHAP / LIME
  • MissForest
  • React
  • Django REST
  • PostgreSQL
  • Ollama
  • Docker
LocalFold: landing page LocalFold: 3D viewer and assistant LocalFold: AI analysis LocalFold: structure comparison LocalFold: demo Impacthon 2026: awards ceremony (podium) at ETSE/USC Impacthon 2026: participant badge (GDG Santiago de Compostela · ETSE/USC)

04 / 2026

LocalFold

Brings protein structure prediction to researchers without HPC expertise: they enter a sequence or a natural-language description and the platform runs AlphaFold2 on CESGA's FT3 supercomputer without touching SLURM, with a Mol* 3D viewer, biological metrics and differential analysis. React 19 and FastAPI on Cloud Run and Firestore, analysis with Gemini/Vertex AI and an A2A multi-agent system (orchestrator and four agents with Google Search Grounding). Challenge from the CAMELIA chair (PLEXUS and USC); award at Impacthon 2026.

  • AlphaFold2
  • React
  • FastAPI
  • Vertex AI
GitHub
MXNJ Outfit Intelligence: matched products MXNJ Outfit Intelligence: outfit catalogue MXNJ Outfit Intelligence: landing page MXNJ Outfit Intelligence: demo HackUDC 2026: award-winning team on stage HackUDC 2026: building the project during the hackathon HackUDC 2026: with the team during the event HackUDC 2026: team trophies, including the Best Use of FOSS AI award

03 / 2026

MXNJ Outfit Intelligence

Fashion visual search: from a photo, it identifies every garment and finds the equivalent product in a 27,000-item Zara catalogue. YOLOv8x-seg anatomical segmentation and an ensemble of SigLIP SO400M and DINOv2 ViT-L/14 that turns each garment into visual embeddings; the catalogue is pre-indexed and retrieval is a cosine-similarity vector search (FAISS) over the normalized vectors. Contrastive fine-tuning (InfoNCE and Triplet) and semi-supervised B2B expansion: 69.66% Recall@15, fully open source. Double prize at HackUDC 2026.

  • SigLIP
  • DINOv2
  • YOLOv8x
  • PyTorch
GitHub
MXP AutoScan: damage detection on the vehicle Datathon award ceremony Datathon first prize Datathon O Camiño dos Datos session

01 / 2026

MXP AutoScan

Automates vehicle appraisal for insurers and workshops: from photos of the car, a Dual Mask R-CNN localises each dent, identifies the affected part and estimates its severity, then generates a structured technical report through a web interface or a REST API. It shortens appraisal time and makes the process more transparent. First prize at the Datathon O Camiño dos Datos.

  • Dual Mask R-CNN
  • Python
  • REST API
News
Spotify Million Playlist Dataset

01 / 2026

Spotify Million Playlist Challenge

Automatic playlist continuation on the 1,000,000-playlist dataset of Spotify's RecSys Challenge 2018: six collaborative-filtering recommenders built from scratch over sparse matrices (user/item KNN, PureSVD, SLIM and FISM), with SLIM and FISM trained via TensorFlow automatic differentiation. A resumable four-stage pipeline evaluated with R-Precision, NDCG and Clicks.

  • Collaborative filtering
  • SLIM
  • PureSVD
  • TensorFlow
GitHub
Federated: FedAvg vs FedProx accuracy Non-IID client label distributions Adaptive streaming models Loss comparison between strategies

11 / 2025

Federated and Streaming Learning

Machine learning when the data cannot move. Federated training over Non-IID clients with Flower and PyTorch, comparing FedAvg against FedProx on MLP, CNN, ResNet-18 and MobileNetV3; and online learning under concept drift with River, with ADWIN drift detection and adaptive models (Hoeffding Adaptive Tree, Adaptive Random Forest).

  • Flower
  • PyTorch
  • River
  • FedProx
  • ADWIN
GitHub
Fuel Price
Prediction

10 / 2025

Fuel Price Prediction

Daily fuel-price forecasting for the gas stations of A Coruña over 1.6M records: a full Spark ML pipeline with temporal feature engineering (lags, rolling means and cyclic encodings), leakage-free splits and Optuna hyperparameter search. Linear Regression, Random Forest and GBT for gasoline, diesel and LPG.

  • PySpark
  • Spark ML
  • Optuna
  • Feature engineering
GitHub
Road segmentation: prediction against ground truth Colour space analysis Best classical result without deep learning U-Net architecture U-Net prediction U-Net prediction U-Net prediction

06 / 2025

Road segmentation

Classical feature engineering and a U-Net over aerial imagery from the Massachusetts Roads Dataset, with a 17-channel interpretable volume per image.

  • U-Net
  • TensorFlow
  • OpenCV
GitHub

Awards

  • 2026 First prize, Datathon O Camiño dos Datos (DIHGIGAL, ITG, UDC, IGAPE). Reference
  • 2026 Second prize, INDITEXTECH challenge, HackUDC. Reference
  • 2026 First prize, best use of open source AI, HackUDC. Reference
  • 2026 Second prize, Google Cloud tools, Impacthon. Reference
  • 2021 Best Outreach award, European CanSat final, European Space Agency. Reference
  • 2016 - 2021 First prizes and honourable mentions at national and international science fairs (Festiciencia, Galiciencia, Ciencia en Acción, Maker Faire).

Contact

xoel.maestu [at] gmail.com
GitHub LinkedIn CV