I turn uncertainty into guidance — and ship it as systems people can use.
DiceTamer is a personal portfolio of data projects, each one born from a problem in my own day-to-day: what a used car is really worth, which skills the job market is asking for, how to keep years of medical results in one place.
Each project is also my excuse to try new technology — in-browser inference, local LLMs, real-time object tracking — and push it far enough to actually use, not just demo.
An end-to-end pipeline that collects second-hand listings from multiple marketplaces, normalizes noisy specifications, and trains a regression stack tuned for outlier resistance — exposed through a Streamlit interface for instant valuation checks.
A privacy-focused upscaler that runs inference entirely client-side, so image files never leave the browser. It streams tiles through ONNX Runtime Web to keep memory predictable while preserving print-grade detail.
A self-hosted AI platform that gives every teacher back their evenings. From a single curriculum it writes homework tuned to each student's level — the hours of hand-differentiating worksheets, done in minutes. The work stays on paper where classrooms already live: sheets are printed, answered by hand, scanned back in, and graded against the teacher's own rubrics — with the teacher steering the AI at every step, never replaced by it.
A conversational assistant that orchestrates pharmacy workflows through retrieval-augmented prompts and guarded tool calls. It answers stock questions, drafts orders, and escalates edge cases to human staff.
A co-authored research project — the final work for the UPC Postgraduate Course in Artificial Intelligence with Deep Learning. It is an end-to-end computer-vision pipeline that turns a raw football match clip into tactical data: detecting players, referees and the ball, holding each identity through occlusion, classifying teams with no manual labelling, and projecting everyone onto a 2D pitch radar. Along the way it benchmarks detector families — YOLO against Faster R-CNN — and trackers from SORT to BoT-SORT.
A deep-learning system that detects players and ball trajectories frame by frame, reconstructs match phases, and outputs tactical heatmaps for coaching review — built to stay reliable under occlusion-heavy conditions.
A browser-based parliament simulator where 25+ historical philosophers — Aristotle, Machiavelli, Marx, Turing, Curie, Nietzsche and more — debate and vote on the decrees you draft, each persona driven by an LLM (local or cloud) grounded in their documented worldview. What began as a menu-driven simulator grew into a full 16-bit JRPG experience: a walkable overworld, branching NPC dialogue, battle-styled debate mechanics, and persistent local saves.
A tactical reference for War Thunder air combat. Pick the fighter you are flying and it works out every enemy you can realistically meet in your battle-rating bracket, then tells you plainly what each one does better than you and what you do better than it, so you know how to fight it before the merge.
Reads SEC annual filings and turns them into a queryable map of what each company depends on — commodities, currencies, countries, suppliers and regulation — next to the audited figures the same companies filed. A deterministic stage turns every filing into traceable Markdown for free; one LLM pass, at about four cents a document, extracts the exposures with a verbatim quote behind each. The result answers questions like “if oil rises, who suffers?” from what companies actually wrote.
An observatory of the data-science job market across Barcelona, Madrid and remote roles: which skills, clouds, languages and seniority levels are in demand, and how that shifts month to month. A local LLM reads every posting, a curated taxonomy folds synonyms together, and a layered PostgreSQL warehouse keeps dated, immutable readings of the market — so every trend can say what share of the corpus it stands on and how it looked at the time. On top, a match layer ranks every posting against my own profile: rules first, then a small decision model reading the full text.
A personal health record that never leaves your computer — because the only way to keep paying taxes is to stay alive. Drop in lab reports and medical PDFs, a local model proposes what each one says, and nothing enters the record until you confirm it beside the original. The payoff is a one-page summary to take to the doctor, built on plain files you can still open in fifteen years.