Work

What I have done, told by what came out of it.

No client names where they do not belong and no internal data. Just the problem, what I did and what was left working.

Pivotal · 2026–today

Governance framework and AI agents for more than 200 users

Problem: The company had AI licences scattered around and no control: nobody knew which tools were in use or with what data, and every department was asking for something different.
What I did: I designed a register of systems, a risk-tier classification and a light approval process. In parallel I built the agents the teams use today, on Anthropic, OpenAI and LangChain, and I own the AI product roadmap.
Result: Usage became mapped and approved, teams stopped working with tools nobody was watching, and the agents made it into the daily work of every department.

IHI · UPM partner

Problem: A European consortium with clinical, academic and industrial partners wanted to anticipate non-adherence. The data sat across medical records, primary care cohorts and questionnaires that did not talk to each other.
What I did: I integrated and analysed more than fifteen healthcare databases, trained supervised models of adherence and side effects, and worked the causal side with Bayesian networks, so we would know not only who drops out but why. Partner coordination and deliverable writing.
Result: Results published in indexed journals and IEEE conferences, three research awards across SEMERGEN and VIATRIS, semFYC and IEEE-EMBS, and a visualization platform evaluated with practitioners before it was called done.

IHI · Horizon Europe · H2020

Problem: Research groups and SMEs have the scientific idea, but the technical section of the proposal gets written by whoever can, at the last minute.
What I did: Writing the AI and data component of more than ten proposals to European calls, with the technical analysis of more than thirty AI systems to hold up the state of the art.
Result: Proposals submitted with a technical part consistent with what has to be executed afterwards, rather than a wish list.

Own research · NEJM AI

Problem: A conversational assistant in production, used by real patients, with no published security review.
What I did: Technical analysis of the system, documentation of the findings and responsible disclosure to the service owner before telling anyone else.
Result: The exposure was fixed and the analysis came out in NEJM AI, with the reading of what this means for any organization putting a chatbot in front of a patient.

Own research · 2024–2025

Problem: Sharing real data between centres holds up entire projects: data protection, ethics committees and timelines that do not fit the research calendar. Synthetic data gets sold as the way out, but hardly anyone measures whether it actually works or what risk it carries.
What I did: First I systematically reviewed what the field was already doing. Then I generated synthetic cohorts and evaluated them in the two directions that matter: whether they kept the statistical utility of the original, and what re-identification risk they carried.
Result: The review came out as first author in the International Journal of Medical Informatics, and the generation work on physical activity time series at CBMS 2025. Enough to decide whether synthetic data solves your case or only postpones it.

H2020 · UPM partner

Problem: A European-scale digital health project with pilots in several countries, each with its own systems, formats and access rules. Without a common layer, every pilot would have ended up as a separate project.
What I did: I worked on data integration and on the AI services running on the platform: making sources of different origin combinable, keeping access controlled, and letting a model be deployed on top without rebuilding it for each site.
Result: The architecture was documented and published in IEEE Access as something reusable, not as the bespoke build of one particular pilot.

GATEKEEPER · H2020 · UPM partner

Problem: A watch measures what it measures well, but on its own it says little about someone living with diabetes, advanced cancer or Parkinson’s. The useful signal shows up when you cross it with what the patient reports and with their clinical history.
What I did: I contributed to multimodal deep learning architectures combining wearable signal, questionnaires and clinical data, much of it on GATEKEEPER data: interstitial glucose estimation, lifestyle biomarkers against HbA1c, quality of life in advanced cancer, Parkinson’s progression and arrhythmia markers in type 2 diabetes.
Result: Five papers published across Scientific Reports, IEEE JBHI and EMBC, each validated on real patient data.