Portfolio

Research into NeRFs and 3D Gaussian Splatting as a pipeline for capturing real-world locations as photorealistic assets for LED-volume virtual production. Evaluated camera setups, algorithms (nerfacto, splatfacto, MCMC), and Unreal Engine integration — benchmarked against RealityCapture and Postshot. The work produced a desktop application that runs COLMAP pose estimation followed by user-chosen reconstruction either locally or on Azure, with a browser-based Gaussian splat viewer.

A computer-vision system that measures how much screen time each brand gets in football broadcasts. A YOLO detector finds the perimeter boards and LED signage; a DenseNet201 classifier sorts each into one of 52 brands — coping with small, motion-blurred crops, a 40× class imbalance, and LED ads that rotate weekly. Runs in production on Azure, validated against human annotators.

ARoundTheWorld is a multiplatform collaborative AR geography application built on the cleAR architecture. It was evaluated with 44 students across three schools and represents the final paper of my PhD, validating that the architecture can produce applications that integrate seamlessly into existing school curricula.

cleAR is a modular, interoperable architecture for building multi-user augmented reality applications in education. Designed from the ground up to bridge the gap between AR's potential and its limited classroom adoption, it was the core contribution of my PhD research.

A retrieval-augmented generation system built for a large European bank, replacing manual SharePoint search with a conversational assistant that answers questions and cites the exact document and page where each answer was found.

End-to-end MLOps platform built on Google Cloud for one of the largest media companies in Spain. Took a portfolio of disconnected ML models: churn prediction, article recommendation and several others, and rebuilt them as Vertex AI pipelines with versioning, monitoring, drift detection and CI/CD across the whole lifecycle.

TRACTION was a Horizon 2020 project, coordinated by Vicomtech, that used opera as a vehicle for social inclusion. My work centred on the Co-creation Stage, the real-time distributed performance tool, spanning development, user-requirements gathering, evaluation, and direct engagement with artistic and community partners across Barcelona, Leiria and Ireland.

A multi-modal system that measures audience engagement at live events by fusing computer vision with WiFi/Bluetooth signal analysis. Designed, implemented, and validated at Vicomtech in 2020, and released as open-source software.

End-to-end real-time analytics for live UFC events: stereo computer vision in the truss above the octagon, accelerometers in the gloves, GPU inference at the venue, and statistics streamed to fans worldwide. Built at AGT International in 2017 and demoed live by our CEO during Werner Vogels' keynote at AWS re:Invent 2017.

Two years as a scientific software developer in Gilles Laurent's department at the Max Planck Institute for Brain Research in Frankfurt, building MATLAB analysis pipelines and R/Shiny web apps for electrophysiologists studying sleep in bearded dragons and visual processing in ex-vivo turtle brains.

An EU FP7 project on indexing, searching and linking large archives of lectures, meetings and video-conferences as interconnected hyper-events. At Fraunhofer HHI I led the video-analysis side of the work — automatic segmentation, slide-transition detection and face-based features for multi-modal speaker linking — published at VISAPP 2014, Interspeech 2014, and the ACM MM 2013 Grand Challenge.

AVATecH was a joint Fraunhofer / Max Planck project that brought state-of-the-art audio and video pattern recognition into ELAN, the de-facto annotation tool used by linguists, anthropologists, and psychologists worldwide, turning weeks of manual labelling into minutes of supervised review.