📂 Projects

Developing an AI-driven system to automatically uplift legacy VR/3D simulations for web accessibility. The system analyzes legacy code and assets to inject ARIA readout text, descriptive tooltips, smart autocompletes, and keyboard-navigable paths. This ensures hundreds of specialized science simulations meet modern accessibility standards without requiring manual per-simulation refactoring.

Labster produces VR/3D interactive science simulations for universities and high schools — immersive and high-quality, but expensive to build. I redesigned the internal course-creation process end-to-end using AI and low-code automation: Airtable as the structured data backbone, n8n as the workflow engine, and Claude + OpenAI as the AI layer. The result cut course development time by 90%, making each simulation dramatically faster and cheaper to produce without sacrificing educational quality. I also built tailored AI-assisted workflows for Learning Experts and Subject Matter Experts — covering learning objectives, driving questions, Bloom's taxonomy levels, and Simpson's skill levels — bridging cutting-edge AI with sound pedagogical frameworks.

Designed and trained a PyTorch-based deep learning model for video movement prediction — from architecture design and dataset preparation through to training and evaluation. (NDA — further details confidential.)

Designed an MVP concept for safe, scalable internal AI adoption at one of the world's largest financial institutions. Researched and synthesized how leading global companies govern AI internally — policies, tooling, access control, audit trails — and translated those patterns into an actionable, enterprise-ready proposal tailored to Banco Santander's specific constraints and culture.

Developed an AI system to estimate property costs by analyzing interior photos. Used image embeddings to train custom Machine Learning models that automatically rated the state of specific rooms (from 0 to 1). The system compared these ratings against an internal valuation database to provide objective, data-driven cost estimates.

Built an end-to-end AI pipeline for UVE Valoraciones that ingested legal and technical property documents (contracts, cadastral records, technical reports) and extracted structured data into a relational database. The system replaced manual expert review, enabling fast and consistent real estate due diligence at scale.

Participated in Helsinki International Junction 2023 among 1,000 international participants. Developed an internal knowledge graph visualization system (similar to Obsidian) using a RAG (Retrieval-Augmented Generation) pipeline for intelligent data exploration.

Winner of the Local Málaga Junction Hackathon (earning a spot for Helsinki International Junction). Developed a steganographic system for Examfy to insert hidden tracking codes into school images. This allowed identifying the source of any leaked materials by tracing back the personal copy to the original recipient.

Developed a real-time AI basketball match evaluation tool for The Embassy, the academy run by KM Calderón — Spanish NBA champion and former NBA player. The system used computer vision to track players and the ball live on court, feeding a dashboard of performance analytics: shot detection, player positioning heatmaps, and match statistics. Designed to run in real time during training sessions with no manual tagging required.

Designed and delivered a comprehensive MLOps curriculum. Covers the full lifecycle: experiment tracking with MLflow, containerisation with Docker, CI/CD pipelines for ML, and model serving. Includes a hands-on "cookbook" repository used by students to master production-grade Machine Learning workflows.

Core curriculum for Machine Learning fundamentals. From Python basics and data manipulation with Pandas/NumPy to building and evaluating deep learning models. This course serves as the technical foundation for bootcamp students entering the AI field.

Six-month intensive sprint: built the initial AI system from scratch and launched a fully working prototype for 3D computer vision padel performance analysis. PADMI went from zero to a demo-ready product using multi-camera 3D tracking, size estimation, and biomechanical metrics. The system analysed player technique, shot speed, and court coverage in real time. Responsible for the full AI stack: camera calibration, stereo reconstruction, model training, and the analytics pipeline.

Developed a cashierless retail store MVP for Mercadona Tech, Spain's largest supermarket chain. The system used 3D computer vision and AI to track customers and products throughout the store — detecting what items are picked up or put back, and automating the checkout process entirely. Key challenges included multi-person tracking under occlusion, real-time inference on in-store hardware, and handling the combinatorial complexity of a full product catalogue.

End-to-end facial expression recognition pipeline: face keypoint detection for data acquisition, geometric normalization to suppress head rotation, manual labeling of six expression classes (normal, angry, happy, sad, surprised, winking), k-NN classifier training, and live webcam inference. Deployed as an interactive Streamlit app — anyone in the audience could test it live during the talk.

Implemented real-time aerial object detection and tracking (people and vehicles) using YOLO and SSD-MobileNet. Deployed these models on both GPU servers and resource-constrained embedded systems on-board drones, ensuring low-latency performance for critical surveillance and monitoring tasks.

Built a thermal anomaly detection system for autonomous photovoltaic panel inspection using drone imagery. Developed Computer Vision models to identify hotspots and malfunctions on solar farms, significantly reducing inspection time compared to manual methods.

Custom YOLOv3 object detection model trained from scratch to recognise a specific cat (Mocka 😂). Used the ImageAI library inside a Jupyter Notebook, covering the full pipeline: collecting hundreds of photos, annotating bounding boxes, training, and validating with the mAP metric. Presented live at TabularConf 2021, my first national conference.

Developed a proprietary pipeline for generating high-resolution multispectral GeoTIFFs from multiple aerial drone cameras. This enabled agronomists to assess crop health via wavelength orthomosaics. Built using C++, Docker, and ROS to handle large-scale image stitching and geographic coordinate alignment.

Master's thesis project (graduated with honors): developed a 3D MRI super-resolution model using Generative Adversarial Networks (WGAN-GP), achieving a 4× improvement in image quality. The model learned to reconstruct high-resolution volumetric MRI scans from low-resolution inputs, with potential applications in reducing scan time and improving diagnostic quality in resource-constrained settings.

Co-founded Monkeat, a real-time marketplace helping pubs boost customer acquisition through live offers during off-peak hours. As CTO, designed the full system architecture and built the Progressive Web App from scratch using Vue + Ionic on the frontend, MongoDB for the database, Flask for the backend API, and Firebase for real-time updates and authentication. Developed inside the Demium incubation program. We could not ultimately validate the MVP and closed the project after a few months — but shipping a full-stack product as a first-time CTO was an invaluable experience.

Part of the Smart-Campus I research projects at the University of Malaga. Built an embedded computer vision system that attached to traffic signs and detected pedestrians in real time, triggering audio and visual warnings to alert drivers — reducing the risk of car-pedestrian collisions. Used YOLO, SSD-MobileNet, and Mask-RCNN running on a Jetson Nano, with Hungarian Algorithm-based object tracking. Research published at CIPI 2019.

Developed a smart signage and analytics system for the University of Malaga campus. Built a Firebase API to log and visualize pedestrian crossing statistics and interaction data. Part of the larger Smart-Campus initiative focusing on urban mobility and data-driven infrastructure.

Co-founded the Málaga chapter of the Spain AI Association — a community for sharing knowledge about Artificial Intelligence through talks, workshops, and courses. Organised and delivered multiple events including bootcamps, meetups, and online courses that reached up to 200 students per session. The association is still active and continues to grow the local AI community.