Research Engineer
Antonio Lobo Santos
LLM Post-training · Interpretability
I work at DFKI in Saarbrücken on the SOOFI project, as part of the post-training team. My focus is interpretability of new LLM architectures.
Saarbrücken, Germany
Research & Engineering Focus
Post-training & Interpretability
Post-training open LLMs at DFKI, and studying what new architectures learn and how their internals change during post-training.
LLM Evaluation
Evaluation pipelines for classifiers and LLMs: synthetic data, custom metrics and moral-capability tests, run on HPC clusters.
LLM Serving & HRI
Low-latency vLLM serving with quantised open-weight models, connected to ROS2 robots for human-robot interaction studies.
Production Software
Two years building Python/TypeScript microservices, Kubernetes deployments and Kafka/NiFi pipelines for monitoring platforms.
About
I'm a research engineer at DFKI in Saarbrücken, where I work on the SOOFI project as part of the post-training team. My focus is interpretability of new LLM architectures.
Before that I was an academic visitor at Oxford, working on learning stability and equilibria in multi-agent systems. At CSIC-IIIA in Barcelona I evaluated the moral capabilities of classifiers and LLMs, and built the LLM serving stack for a socially assistive robot.
I started out as a software engineer at Axión in Seville, working full-time while studying Mathematics and Computer Engineering.
Research & Engineering Experience
Research Engineer - LLM Post-training & Interpretability
DFKI, German Research Center for Artificial Intelligence
Member of the post-training team in the SOOFI project, which builds open foundation models. I work on interpretability of new LLM architectures.
Academic Visitor - Multi-Agent Systems & Game Theory
University of Oxford
Worked on game-theoretic conditions for learning stability and equilibria in multi-agent systems.
Research Intern → AI Research Engineer - LLMs, Robotics & Evaluation
CSIC-IIIA, Spanish National Research Council
- Built moral-capability evaluations for text classifiers and LLMs, based on Moral Foundations Theory and Schwartz's theory of values.
- Generated synthetic datasets and designed metrics for robustness and value-alignment tests.
- Fine-tuned BERT-style classifiers with custom losses; ran experiments on HTCondor and multi-GPU servers.
- Built the vLLM serving pipeline (quantised open-weight models) for real-time HRI studies.
- Integrated the dialogue components with the ROS2 robot stack, including ASR/TTS.
Lecturer in Probabilistic Graphical Models
University Pompeu Fabra
Taught probabilistic graphical models, with a focus on Bayesian inference.
Undergraduate Researcher - LLMs & Mathematical Knowledge Graphs
University of Seville
Built a LangChain/RAG pipeline that extracts mathematical knowledge from LaTeX into OWL knowledge graphs. This became my bachelor's thesis and a journal paper.
Software Engineer - Platform, Microservices & Real-Time Monitoring
Axión
Full-time while completing dual B.Sc. degrees in Computer Engineering and Mathematics.
- Built Python/FastAPI and TypeScript/NestJS microservices for monitoring platforms used in transport, ports and public spaces.
- Set up an on-prem high-availability MicroK8s cluster and migrated the company's web apps to it with GitLab CI/CD.
- Ran Kafka and NiFi ETL pipelines; built real-time features with WebSockets, GraphQL and Elasticsearch.
- In the second year, led a small team and mentored two interns.
Selected Projects
MariChatmen - written Andalûh language model
Qwen3.5 adapted to written Andalusian Spanish: tokeniser expansion, continued pretraining, SFT and ORPO.
Moral-capability evaluation for LLMs and classifiers
Evaluation of moral reasoning in BERT-style classifiers and LLMs, using synthetic data and custom metrics.
Real-time LLM serving for HRI
vLLM serving of quantised open-weight models for a ROS2 robot, with structured dialogue and ASR/TTS.
Deep Open-Set Recognition framework
PyTorch Lightning/Hydra framework for open-set recognition with ResNet, EfficientNet, ViT and DINOv2 backbones.
View RepositoryPPO/GRPO continuous-control RL framework
PPO and critic-free GRPO implemented from scratch for continuous control.
View RepositoryFlood-IDSS
Flood decision-support app combining data sources, ML models, an API and dashboards.
View RepositoryMathematical Knowledge Graphs with LLMs
LLM/RAG pipeline that turns LaTeX mathematics into OWL knowledge graphs.
Read PublicationAB Data Challenge - Finalist
Transformer-based anomaly detection on utility time-series data.
ShadowPals
Shadow-puppetry learning app using MediaPipe hand landmarks; I wrote the hand-pose parametrisation.
View RepositoryEarlier Projects & Awards
App Inventor Against Cyberbullying
Mobile application recognised as MIT App Inventor "App of the Month" for community cyberbullying awareness.
Read ArticleSmarthuerto
Award-winning smart agriculture app that won the EC2CE Grow-Lab contest and was featured by MIT App Inventor.
Read ArticlePublications & Talks
Publications
EMY: Supporting Autism Therapy with a Socially Assistive Robot
HRI '26: 21st ACM/IEEE International Conference on Human-Robot Interaction
I worked on the LLM dialogue and system integration.
View ArticleEnhancing Mathematical Knowledge Graphs with Large Language Models
Modelling (MDPI), based on my bachelor's thesis
View ArticleTrustworthy AI Through Dual-Role Reasoning
Dual-role reasoning as a way to evaluate and improve the safety of LLM reasoning.
Current Research
Interpretability of new LLM architectures
At DFKI, within the SOOFI post-training team.
Moral-capability evaluation for classifiers and LLMs
Manuscript in preparation.
NLP pragmatics and formal grammars
Pragmatics problems studied with context-free grammar formalisms.
Invited Talks
Invited Speaker: LLM-Tools for Coding
La Moncloa (Spanish Government)
How to use LLM coding tools well, and where they go wrong.
Seminar: LLM-Tools for Coding
CSIC-IIIA
Seminar for IIIA researchers on prompting and working with coding assistants.
View Seminar DetailsEarlier Publications
Building the "mapamático" (map-matic)
Epsilon Journal of Mathematics Education, Number 114
View PDFInfinity: Some Curiosities and its Teaching in High School
Números Journal, Volume 113
View ArticleEducation & Training
M.Sc. in Artificial Intelligence
UPC–UB–URV
Honours in Machine Learning, Deep Learning, Reinforcement Learning and Complex Networks.
B.Sc. in Mathematics
University of Seville
Probability, statistics, optimisation and analysis.
B.Sc. in Computer Engineering
University of Seville
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
Read ThesisCertifications
