Selected work
Projects chosen to show how I frame problems, build evidence and decide what is—and is not—ready for real use.
Data engineering · Multilingual NLP
A traceable pipeline that transforms noisy multilingual source data into aligned corpora across Catalan, English, French, Italian and Spanish.
Cut a core workflow from 48 to 6 hours and released the resulting corpus as an open, reproducible data product.
Applied ML · Model risk
A leakage-aware temporal modelling pipeline for prescription refill risk, with calibration and explicit analysis of dataset shift.
Surfaced a material performance drop on the later test period—evidence that the model should not be deployed without addressing temporal shift.
Forecasting · Decision science
A simulation-based, out-of-sample comparison of statistical, structural and machine-learning approaches to foreign-exchange forecasting.
Found that random walks remained difficult to beat at short horizons, while structural hybrids improved longer-horizon forecasts for selected currency pairs.
Applied AI · System design
An experimental retrieval system combining order histories, policy documents and task-specific tools for support questions.
Currently being rebuilt around deterministic business rules, structured data and an explicit evaluation set.
Earlier engineering work
An early robotics project on coordination, mapping and sensor-based navigation for high-risk mining environments, published by Springer.
View publicationEngineering principles
A portfolio should expose judgement—not only polished outcomes.
Clear inputs, versioned code, explicit dependencies and a path from raw data to result.
Baselines, untouched test data and metrics that match the decision the system needs to support.
Runtime, lineage, failure modes and maintainability considered alongside model performance.