Full-stack legal SaaS built on a custom RAG system over 23,000+ legal fragments indexed in Qdrant. The corpus covers the National Constitution, the full Civil & Commercial Code (~2,600 articles), the Criminal Code, labor law (15+ statutes), 127 Supreme Court rulings, collective bargaining agreements and ~7,920 real (anonymized) lawyer queries, plus live economic indicators. Node.js + Qdrant backend, React frontend, authentication, query history, legal blog, severance calculator and AI contract generation. Freemium model in production.

Interactive 3D map (CesiumJS) that cross-references satellite data with territorial information to reveal patterns between wildfires and land-use change over protected native forest in Córdoba (2018–2025). It overlays burn scars derived from Sentinel-2 via Google Earth Engine (dNBR index classified by severity) with near-real-time NASA FIRMS active fires, Open-Meteo winds, native forest, municipal land parcels and historical recurrence. A composite score ranks the 20 scars with the most alert signals. Each scar is analyzed with Claude through a Cloudflare Worker that keeps the API key server-side. Python + GEE data pipeline, Excel/PDF export and auto-updates via GitHub Actions. Runs 100% static on GitHub Pages.

Real-time global seismic visualizer built on CesiumJS: a 3D globe with real satellite imagery, high-performance WebGL rendering and native geographic coordinate support. It consumes up to 5,000 events from the USGS catalog and places them on the globe in real time. Real-time vector winds over 77 grid points from Open-Meteo. Configurable push alerts by magnitude. Dual seismic-risk model: an ETAS implementation in the browser (Aki b-value 1965, Omori-Utsu law, Gutenberg-Richter + Poisson) and an XGBoost model trained in Python on 2 years of catalog data, 10 seismic features and 5-fold cross-validation (AUC ~0.84–0.87). Fully client-side — no backend.

Audio DSP + ML pipeline that extracts 67 acoustic features (pYIN, 13 MFCCs + deltas, HPSS, spectral centroid, onset detection), classifies the vocalization into 9 classes using a Random Forest trained with weak supervision over Freesound, and calls Claude Opus via SSE streaming to produce an interpretation grounded in peer-reviewed literature. Feature extraction in pure numpy/scipy in production to avoid OOM on free tiers.