AI Engineer focused on applied machine learning, AI systems and data-intensive products.

BSc in Artificial Intelligence — Universidade da Coruña (2022–2026). Previously AI Engineer Intern at ABANCA, working on GenAI, RAG reliability and forecasting (9.2/10 final review).

📰 Featured in El Español / Quincemil for my independent ML and data analysis of A Coruña's BiciCoruña bike-sharing network.

I work on AI systems beyond the model itself: where the data comes from, how outputs are evaluated and validated, what the system is allowed to touch, and what happens when it is uncertain or wrong.

Models are components. Reliable systems are the product.

📍 A Coruña, Spain · LinkedIn · Email

Real-time Data Engineering · Machine Learning · Urban Mobility

Independent technical analysis of A Coruña's public bike-sharing network, from building the historical dataset to predictive modeling and operational recommendations.

There was no consolidated public historical dataset for the service, so I built it from scratch.

• Built and deployed a Node.js collector polling the GBFS feed of all 79 stations every 5 minutes
• Developed a LightGBM stockout classifier with a 30-minute prediction horizon
• Achieved F1 = 0.78 and 87% accuracy
• Tuned the operational stockout threshold to provide useful reaction time before complete unavailability
• Segmented stations into 4 behavioral archetypes using K-Means
• Quantified spatial and temporal service failures across the network
• Tested transfer learning from approximately 500k Barcelona Bicing records; it produced no meaningful improvement over the local model, so I discarded the additional complexity and documented the negative result
• Converted the analysis into concrete fleet-rebalancing and capacity recommendations

The project demonstrates the complete chain:

data collection → feature engineering → experimentation → evaluation → operational decision

Python · LightGBM · Pandas · Node.js · Streamlit · Leaflet · Railway

→ Repository

Graph ML · Representation learning · Browser-side inference

A large-scale music intelligence system for exploring artist relationships through collaboration graphs and learned embeddings.

The embeddings were not the hard part. Shipping them was.

• Built a public collaboration graph containing 100,000 artists and 1,059,326 edges
• Constructed and processed the graph using DuckDB
• Trained 64-dimensional Node2Vec-style embeddings from graph random walks
• Projected the latent space into 3D with UMAP
• Built interactive exploration with Three.js
• Added shortest-path discovery between artists with track-aware route explanations
• Quantized embeddings from float32 to int8 for client-side inference
• Compressed the public embedding bundle to approximately 6.1 MB
• Implemented MixDNA lite directly in the browser
• Shipped the complete experience as a static site with no backend dependency

The engineering question became:

How do you turn a large offline ML pipeline into something users can open in a browser and use immediately?

Python · DuckDB · Gensim · UMAP · JavaScript · Three.js

→ Live Demo · Repository

Time-series forecasting · ML experimentation · Model selection

Electricity-demand forecasting system for Spain comparing classical machine learning, deep learning and statistical baselines.

• Built preprocessing and feature-engineering pipelines for temporal data
• Engineered cyclical time features, lag variables, rolling statistics and calendar information
• Compared LightGBM, XGBoost, Temporal Fusion Transformer and statistical baselines
• Best model: LightGBM — 1.15% MAPE
• XGBoost: 1.46% MAPE
• Seasonal-naive baseline: approximately 9.5% MAPE
• Analyzed why the more complex deep-learning experiment underperformed instead of hiding the result
• Structured the project as a reproducible pipeline with separate data, model and evaluation components

The conclusion I care about:

Model complexity has to earn its place.

A simpler model with the right representation of the problem can be substantially more useful than a more sophisticated architecture deployed for its own sake.

Python · LightGBM · XGBoost · PyTorch · Darts · Pandas

→ Repository

Geospatial analytics · Decision systems · Renewable energy

Multi-criteria decision-support system for identifying suitable wind-farm locations in Galicia while accounting for environmental risk.

• Integrated Global Wind Atlas raster data at 100 m height
• Processed 40,000+ historical lightning observations from MeteoGalicia
• Converted discrete lightning events into a continuous spatial-risk surface using Gaussian KDE
• Applied geographic land masks using GeoJSON administrative boundaries
• Combined wind-resource potential and lightning risk into a suitability model
• Validated high-suitability regions against existing wind-turbine locations retrieved through the OpenStreetMap Overpass API
• Built interactive 2D, 3D and temporal geospatial visualizations

This project sits at the intersection of data engineering, geospatial analysis and decision modeling rather than conventional supervised ML.

Python · Rasterio · GeoJSON · Folium · PyDeck · GIS

→ Repository

Product engineering · Urban mobility · Edge-first web architecture

Real-time public transport Progressive Web App built for A Coruña.

I designed and developed the product end-to-end, from transport-data integration and routing logic to the mobile interface and deployment architecture.

• Built a mobile-first PWA with Vanilla JavaScript
• Integrated public-transport data, geolocation and interactive maps
• Used local caching and Cloudflare infrastructure for near-instant route searches
• Designed a privacy-first architecture with client-side processing
• Avoided third-party advertising and tracking dependencies
• Built for fast loading and usability on mobile connections

BUS-CO is useful to me as a portfolio project for a different reason than the ML work: it demonstrates that I can take a technical idea all the way to a user-facing product.

JavaScript · Cloudflare · Leaflet · HTML · CSS · Git

Academic research involving machine-learning methods applied to astronomical data from the Gaia and Euclid missions.

The work focuses on the characterization of stellar systems and the search for substellar companions using spectroscopic and photometric information.

It involves:

• Large-scale scientific data processing
• Feature engineering
• Classification and representation problems
• Neural-network experimentation
• Model evaluation
• Astronomical catalog cross-matching
• Analysis of model limitations and scientific uncertainty

Research carried out in the context of CITIC / LIA2 and my BSc final project.

Business Process Optimization · 2025

Worked on GenAI and predictive systems inside a banking environment.

Final performance review: 9.2/10

• Designed a RAG-Readiness framework to evaluate internal documentation before LLM ingestion
• Built logic to identify documentation issues that could degrade retrieval quality
• Worked on context-management strategies for customer-support tooling
• Reduced hallucination rates by approximately 40% in long-context sessions
• Worked with Gemini-based systems under strict privacy constraints

A major lesson from the work was that many apparent "LLM problems" are actually data, context or system-design problems.

• Developed a LightGBM demand-prediction model for service-volume forecasting
• Designed validation logic that checks whether the available data supports ML inference
• Added fallback behavior to statistical baselines when the predictive signal is insufficient

Rather than always forcing an ML prediction, the system could degrade to something simpler and more defensible.

All work was performed inside self-contained environments under banking security and data-privacy requirements.

Python · LightGBM · LLMs · RAG · Data Analysis

namase.es

Technology studio focused on applying software, automation and AI to concrete business processes.

My work includes:

• Understanding how information moves through a business
• Identifying repetitive or failure-prone processes
• Connecting previously isolated systems
• Designing workflow automations and internal tools
• Defining which cases can be automated and which require human review
• Building AI-enabled systems with explicit validation and fallback paths
• Developing web products and commercial infrastructure

The principle behind the work is simple:

Automation should reduce operational complexity, not hide it behind another layer of software.

BSc in Artificial Intelligence 2022–2026

The degree covered the full stack behind intelligent systems rather than ML in isolation.

Core areas included:

• Machine Learning
• Advanced Machine Learning
• Deep Learning
• Natural Language Processing
• Information Retrieval & Web Mining
• Computer Vision
• Knowledge Representation & Reasoning
• Multi-Agent Systems
• Mathematical Optimization
• Algorithms & Data Structures
• Databases
• Software Engineering
• Concurrent, Parallel & Distributed Computing
• Data Processing & Analysis
• AI Systems Development & Deployment
• Cybersecurity
• AI Law, Security & Ethics

Start with the baseline. I want to know how much value the sophisticated approach actually adds.

Complexity has to justify itself. A more advanced model is not automatically a better system.

Negative results belong in the project. If an experiment does not work, understanding why is part of the result.

Evaluation comes before confidence. A system should know when the available evidence is insufficient to trust its own output.

Failure paths are part of the architecture. I care about what happens when a model is wrong, an input changes or an external dependency fails.

The final artifact should be usable. I prefer systems people can open, test and understand over notebooks that only work in the author's environment.

I'm particularly interested in problems involving:

• AI Engineering
• Machine Learning Engineering
• Applied Machine Learning
• LLM & RAG systems
• Data-intensive products
• Forecasting
• ML evaluation and reliability
• Data pipelines
• Intelligent decision-support systems

Especially when the challenge is not simply training a model, but making the complete system reliable and useful.

📍 A Coruña, Spain

LinkedIn · Email

Open to AI Engineering, Machine Learning Engineering and applied ML opportunities.