NLP and Graph Representation Learning
I design NLP experiments that bring graph structure into language tasks and make evaluation more reproducible.
NLPGraph MLPyTorch
Research engineerJohn Stewart
John Stewart
Fabila-Carrasco
Research Engineer in NLP, Graph ML, and Structured Data
I build machine-learning methods for structured and noisy data, from NLP and graph representation learning to scientific ML and decision-support analytics.
- NLP
- Graph ML
- Scientific ML
- PyTorch + CUDA
- Reproducible experimentation
Current work
NLP + graph representation learning
Previous work
Scalable graph clustering on large graph datasets
Applied delivery
Nationwide analytics, dashboards, and decision support
Record
15 peer-reviewed papers
Communication
20+ talks across research communities
Selected work
Three image-led case studies that show where my work is most useful.
These are the three areas I would point to first if you want the shortest route through my research-engineer profile.
Scientific ML for EEG, fMRI, and Difficult Signals
I adapt graph-signal methods to EEG, fMRI, and other noisy scientific datasets where structure matters.
Scientific MLEEG / fMRIGraph signals
Public-Sector Analytics and Decision Support
I turn large national datasets into reproducible dashboards and analysis that help non-technical stakeholders act.
Decision supportR + ShinyPublic-sector analytics
Experience snapshot
A concise view of the roles that best explain the bridge between research depth and practical delivery.
These four roles are the clearest path through the mix of current NLP work, previous graph methods, scientific ML, and stakeholder-facing analytics.
Sep 2025 - Apr 2027
Postdoctoral Researcher in NLP and Graph Representation Learning
University of Cardiff
Developing NLP methods that combine language modelling with graph reasoning, supported by GPU-accelerated training and evaluation workflows in Python and PyTorch.
AI/MLNLPGraph ML
Feb 2024 - Apr 2025
Postdoctoral Researcher in Scalable Spectral Clustering
University of Edinburgh, School of Informatics
Designed clustering algorithms for large graph-structured datasets, with a focus on computational efficiency and scalable experimentation for high-dimensional network analysis.
Graph MLScalable learningClustering
Nov 2020 - Jan 2024
Postdoctoral Researcher in Graph Signal Processing for Biomedical Data
University of Edinburgh, School of Engineering
Developed nonlinear graph-signal methods for EEG and neuroimaging time series, leading the computational analysis work for a Leverhulme-funded project on difficult healthcare data.
HealthcareGraph Signal ProcessingEEG
Feb 2014 - Sep 2015
Industry and Government Data Science Researcher
National Institute of Educational Evaluation (INEE)
Led end-to-end statistical modelling for nationwide assessment and census data, building reproducible R pipelines and Shiny dashboards for policy and resource-allocation decisions.
Applied analyticsDecision supportDashboards
Selected publications
A small publication set that supports the homepage story without taking it over.
I keep the full record searchable elsewhere, but these are the quickest publication examples for an industry-facing first read.
International Conference on Machine Learning (ICML) • 2025
Signed Laplacians for Constrained Graph Clustering
A graph clustering method that respects constraints while remaining mathematically principled and scalable enough to matter for real structured datasets.
AI/MLGraph MLClustering
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) • 2024
Graph-based Permutation Patterns for the Analysis of Task-related fMRI Signals on DTI Networks in Mild Cognitive Impairment
A way to study how brain activity evolves across anatomical networks rather than treating signals as isolated time series.
HealthcareGraph Signal ProcessingApplications
Chaos, Solitons and Fractals • 2023
Dispersion Entropy for Graph Signals
An entropy measure designed for graph-shaped data, making it easier to quantify complexity in signals that depend on network structure.
Graph Signal ProcessingApplicationsScientific ML
How I work
I try to make difficult technical work measurable, reproducible, and easy to reason about.
This is the working style I try to bring to open-ended research questions, hard data, and mixed technical audiences.
Translate questions into experiments
I turn open-ended research questions into tractable comparisons, baselines, and evaluation plans.
Work well with hard data
I am comfortable with noisy, structured, high-dimensional datasets where simple benchmark assumptions do not hold.
Communicate across audiences
I explain technical work clearly to collaborators in AI, signal processing, mathematics, and policy settings.
Keep workflows reproducible
I build clear experimentation logic so methods, evidence, and trade-offs are easier to inspect and reuse.
I like problems where research depth and practical delivery have to coexist.
I like work that sits between research depth and practical usefulness, especially when language, graphs, scientific ML, and real stakeholder needs all meet in the same problem.
That usually means language, graphs, scientific data, or policy-facing analytics, with enough structure and noise that careful experimentation matters.
I’m happy to hear from research groups, product teams, and collaborators working on language, graphs, or difficult structured data.
I’m especially happy to hear about NLP, graph ML, scientific ML, and stakeholder-facing analytics problems where careful experimentation matters.
Fabila-CarrascoJ@cardiff.ac.uk
LinkedIn and Google Scholar are the best places to browse a compact professional profile and the full academic record.