AI Engineer & Applied Researcher. I build AI systems whose behaviour can be measured, verified and trusted: LLM evaluation and verification, digital twins and calibrated forecasting.
I test AI systems before they go into production. Models, agents and physical simulations: whether they work, where they fail and why.
For twenty years before AI, I was on the other side of that decision. I ran plants, sales operations and P&Ls, and signed the deployments myself. So I test for the person who has to sign: the manager who answers to a board, an auditor or a regulator.
Recent work (client names and references available on request):
- Banking: evaluated a customer-facing retail-banking agent, a WhatsApp support assistant and a model-risk Q&A pipeline. Tested whether grounding calibrated in English still works in Spanish, Catalan and Galician.
- Healthcare: built and tested an LLM pipeline for emergency-department triage.
- Energy: physics-based digital twin for grid-scale battery storage that flagged end of life in 44 of 44 cells, on average 76 cycles in advance. Presented at the IEEE PES General Meeting 2026. Open source: Groundlens (guardrails for LLM answers) and Otwin (a physics engine for industrial assets), both on PyPI.
If your AI system has to go into production and someone has to sign for it, send me a message to javier@jmarin.info
| Paper | Venue | Code |
|---|---|---|
| Geometry of Validity: Analyzing Trajectories Beyond Linear Probes | NeurReps, NeurIPS 2026 workshop | code |
| Physics-informed digital twin for grid-scale energy storage | IEEE PES General Meeting 2026, invited talk | otwin-hybrid |
| Hamiltonian Neural Networks for Robust Out-of-Time Credit Scoring | IEEE DSA 2025 | code |
| Semantic Grounding Index: Geometric Bounds on Context Engagement in RAG Systems | arXiv, 2025 | groundlens |
| Perplexity-Based Adjacency for Uncertainty Quantification | arXiv | code |
Full list of papers: Google Scholar