Behavioural data scientist — I build data systems that show their work, so a complex and opaque reality becomes legible for the people who actually have to act on it.

What I care about isn't the model — it's whether the output can be trusted: does it cite its source, does it know how sure it is, and does it fail loudly instead of overselling. I work at the intersection of behavioural science, rigorous measurement, and dependable data engineering.

A note on AI, since everyone claims it: in my projects the model is rarely more than a thin layer — deliberately. Every decision and every alert is computed by deterministic code; the model handles background memory and summarisation. The work I actually do is the unglamorous part underneath: pipelines, measurement, provenance, and evaluating when a system is genuinely ready rather than good-on-average.

I hold a postgraduate Master's in Behavioural Data Science (IL3 – Universitat de Barcelona, 9.17/10). Before specialising, I spent nearly three years as a Business Analyst on Santander's Confirming platform, as the pre-production QA owner — where I learned institutional rigour by being the last check before a rollout that would have miscalculated credit limits for thousands of suppliers.

• Political-data observatory — where has the vote for Vox grown in Spain between elections, and does the answer change with the territorial scale you measure at? Official Ministry of the Interior results, published down to polling-station level, tested across four levels: polling station → municipality → province → autonomous community. An earlier cross-country design was executed, found not viable with its sources, and retired — documented rather than hidden. (My main line of work.)
• Acceptance infrastructure for AI agents — AI agents write code in my projects overnight, unattended. Every task carries the command that decides whether it is done, and an agent cannot commit until that command passes. More than 160 automated checkers across four of my systems, and the harness breaks its own checkers on purpose to find the ones that fail to notice. "Could not measure" is a verdict of its own, never counted as a pass. (Private — happy to walk through it.)
• Retrieval systems you can trust — a retrieval (RAG) system with its own evaluation harness and a deterministic regression gate (it hard-fails if retrieval quality drops), so the system catches its own regressions instead of finding out in production. (Private — happy to walk through it.)
• Responsible AI in real apps — an endurance training and nutrition platform built end-to-end (FastAPI, PostgreSQL, Supabase, Garmin/Strava, over 2,500 automated tests). Every decision and every alert that matters is computed by deterministic code. The model does background memory and summarisation, and nothing that decides anything. (Private — available on request.)
• Geospatial for decisions — a dashboard built for a UN ESCWA assignment over Lebanon (Leaflet.js, 1,545 ADM3 localities), with a rule-based demographic classifier and prompt-level defenses so answers stay grounded in the data.
• Making data legible — end-to-end analytics for Project RYSE in R (clustering, Random Forest, XGBoost, GLM, ETL), surfacing a decision gap rather than a skill one; and a World Happiness Streamlit dashboard for cross-country wellbeing.

Python · R · SQL · FastAPI · PostgreSQL · Leaflet.js / GIS · Streamlit · R Shiny · Git

Roles where data has to be trustworthy to matter — reliable systems, honest measurement, provenance without overselling causality. That spans information integrity & accountability, evidence & impact, and behavioural insights, in teams that publish their methods and welcome scrutiny. Fully mobile across the EU.