I build evaluation, inference and software systems for difficult AI problems - from LLM safety evaluation and real-time human-robot interaction to production data platforms and on-prem infrastructure.
Oxford, UK / Seville, Spain
Designing evaluation pipelines for LLM robustness, moral-capability analysis and failure-mode discovery. Experience with synthetic data generation, custom metrics, adversarial prompting, value-alignment tests and computational experiment execution on shared infrastructure.
Building low-latency LLM serving pipelines with vLLM, quantisation and open-weight models for multi-user research environments and real-time HRI experiments.
Integrating LLM-side interaction modules into ROS2-based socially assistive robotics systems, with structured dialogue, ASR/TTS coordination and real-time constraints for lab-based HRI studies.
Training and evaluating transformer-based and BERT-style models using shared HPC, Slurm/HTCondor workflows and multi-GPU servers, with reproducible experiment configuration and analysis pipelines.
Research-level exposure to reinforcement learning, including from-scratch PPO and no-critic GRPO implementations for continuous-control experiments, plus coursework honours in Reinforcement Learning.
Designing production Python/TypeScript microservices, APIs, CI/CD pipelines, Kubernetes/MicroK8s deployments and Kafka/NiFi data-processing systems for real-time monitoring platforms.
I am a research engineer and ML systems builder working across LLM evaluation, efficient model serving, AI safety, robotics/HRI and production data platforms. I enjoy problems where scientific uncertainty meets engineering constraints: systems that need rigorous modelling, careful evaluation and reliable software.
At CSIC-IIIA, I design evaluation workflows for moral capabilities, robustness and model behaviour in text classifiers and LLMs. My work includes synthetic data generation, custom metric design, BERT-style classifier fine-tuning, open-weight LLM evaluation and computational experiments on shared HPC and multi-GPU infrastructure.
I also build efficient LLM-serving pipelines for real-time human-robot interaction, using vLLM, quantisation and structured dialogue components integrated with ROS2-based robotic systems. The goal is practical AI behaviour under real interaction constraints, not just offline benchmark performance.
Before AI research, I worked full-time as a software engineer at Axión, building production microservices, on-prem Kubernetes/MicroK8s infrastructure, GitLab CI/CD, Kafka/NiFi ETL pipelines and real-time monitoring systems for public-service and private infrastructure.
At Oxford, my current work focuses on game-theoretic results for learning stability and equilibrium in multi-agent systems, with applications to multi-agent coordination and learning dynamics.
Built the LLM-side serving and integration stack for ROS2-based HRI lab studies: vLLM serving, quantised open-weight models, structured dialogue components, ASR/TTS coordination and real-time inference constraints.
Designed evaluation workflows for moral-capability analysis in text classifiers and LLMs, grounded in moral psychology theories such as Moral Foundations Theory and Schwartz's theory of values. Built synthetic datasets, custom metrics and experiment pipelines for classifier and LLM evaluation.
Main developer on production monitoring systems for public-service and private infrastructure. Built Python/FastAPI and TypeScript/NestJS microservices, APIs and real-time monitoring components; operated Kafka/NiFi ETL workflows; deployed GitLab CI/CD; and migrated web applications to an on-prem high-availability MicroK8s/Kubernetes platform.
Co-designed and developed a modular PyTorch Lightning framework for open-set recognition experiments, with configurable backbones, OSR strategies, Hydra experiment configuration, standardised metrics and visualisation tools.
Integrated the components of a flood decision-support project into a full-stack application combining data integration, ML models, APIs and interactive dashboards for flood risk analysis and decision support.
Developed the mathematical parametrisation layer for an interactive shadow-puppetry learning platform using computer vision, MediaPipe hand landmarks and real-time feedback.
Studying game-theoretic conditions for learning stability and equilibrium in multi-agent systems, with applications to multi-agent coordination and learning dynamics.
Designed and delivered lectures on probabilistic graphical models for AI, covering theoretical foundations and practical applications, with emphasis on efficient Bayesian inference in structured probabilistic models.
Built and evaluated a LangChain/RAG pipeline for extracting mathematical knowledge from LaTeX sources into OWL knowledge graphs. The work became my bachelor's thesis, was selected as a university finalist and later resulted in a peer-reviewed journal publication.
Full-time while completing dual B.Sc. degrees in Computer Engineering and Mathematics.
Experimental Qwen3.5 adaptation for written Andalusian Spanish, covering tokeniser expansion, Andalûh adaptive pretraining, SFT, ORPO, persona tuning and release-gate evaluation.
Synthetic data, custom metrics, moral psychology, BERT-style classifiers, LLM evaluation, HPC/HTCondor and multi-GPU experiments.
vLLM, quantised open-weight models, ROS2, structured dialogue, ASR/TTS coordination and real-time lab-study constraints.
PyTorch Lightning/Hydra framework with ResNet, EfficientNet, ViT and DINOv2 backbones, OSR metrics and visualisation.
View RepositoryFrom-scratch course project implementing PPO and no-critic GRPO variants, configurable experiments, logging, checkpoints and performance comparison.
View RepositoryFull-stack flood decision-support application integrating data sources, ML models, backend APIs and interactive dashboards.
View RepositoryLLM/RAG pipeline for extracting mathematical knowledge from LaTeX into OWL knowledge graphs; resulted in peer-reviewed publication.
Read PublicationTransformer-based time-series anomaly detection for utility data; designed training code and deployed experiments on distributed hardware.
Computer-vision learning platform using MediaPipe landmarks and mathematical hand-pose parametrisation.
View RepositoryMobile application recognised as MIT App Inventor "App of the Month" for community cyberbullying awareness.
Read ArticleAward-winning smart agriculture app that won the EC2CE Grow-Lab contest and was featured by MIT App Inventor.
Read ArticleHRI '26: 21st ACM/IEEE International Conference on Human-Robot Interaction
Contributed to LLM-supported interaction and system integration for a socially assistive robotics system.
View ArticlePeer-reviewed journal article derived from my bachelor's thesis
Built and evaluated an LLM-assisted pipeline for extracting mathematical knowledge from LaTeX into queryable knowledge graphs.
View ArticleSafety-oriented reasoning and evaluation work for LLM systems
Explored dual-role reasoning as a safety-oriented approach to evaluating and improving reasoning behaviour in LLM systems.
Designing evaluation workflows grounded in moral psychology, with synthetic data generation, custom metrics and HPC/multi-GPU experiments.
Researching NLP problems involving pragmatics and classical context-free grammar formalisms.
La Moncloa (Spanish Government)
Delivered an invited seminar on effective and proper use of LLM tools for coding, focused on best practices, capabilities and safety considerations for software engineering workflows.
CSIC-IIIA
Conducted a technical seminar for researchers at IIIA on advanced usage of LLM tools for coding, with emphasis on prompting strategies, assistant workflows and safe implementation practices.
View Seminar DetailsEpsilon Journal of Mathematics Education, Number 114
View PDFNúmeros Journal, Volume 113
View Article
Focus: reinforcement learning, deep learning, statistical modelling, ML systems and AI research. Honours in Machine Learning, Deep Learning, Reinforcement Learning and Complex Networks. Selected projects include from-scratch PPO/GRPO implementations, open-set recognition, flood decision support and computer-vision interaction systems.
Strong training in probability theory, statistics, optimisation, mathematical modelling, and rigorous analysis. Developed a deep theoretical foundation for advanced machine learning and data science.
Graduated with 9 honour distinctions, including Algorithms and Complexity, Computer Architecture, Computer Networks, and Intelligent Systems.
Bachelor's Thesis: Large Language Models and Knowledge Graphs to Model Mathematical Knowledge
Developed and evaluated an LLM-based pipeline for extracting mathematical knowledge from LaTeX into a structured knowledge graph.
Read Thesis
Working languages for research, engineering and collaboration.