Explore/exploit benchmark for LLMs. Runs models through 8 environments — stationary and non-stationary bandits, grid search, combinatorial optimization — and measures cumulative regret, exploration ratio, and adaptation speed. Useful for comparing behavior across models, temperatures, and prompt framings.
AI enrichment pipeline for debt collection cases. Takes a minimal case record (name, country, debt amount) and fans out across 23 OSINT modules to produce a sourced, collector-ready dossier. Uses a wave-based async scheduler and a typed Signal system for provenance tracking.
Telegram bot for managing a shared home pantry. Add, remove, list items, and get recipe suggestions via slash commands, free-form messages, or photos of groceries or receipts. Uses an LLM for natural-language parsing.
A daily Telegram bot that parses the RSS feeds from Spain's
BOE and Galicia's
DOG, filters entries relevant to my interests with AI, and sends a summary to Telegram.
App for managing
Beancount ledgers, built with NiceGUI. Point it at a
.beancount file to visualize asset allocation, stock and crypto holdings, transactions, document attachments, and a bank import flow via the Enable Banking API with automatic payee categorization.
A skill for coding agents (claude code, opencode...) for adding i18n translations to big codebases that originally had hardcoded strings.
Competed on a tree canopy detection challenge using satellite imagery. Placed top 50 out of 4000+ participants. Explored Yolo, EOMT (DinoV3 backbone) and Mask2Former architectures.
code repository.
We released a fanzine with some of our favorite articles from our blog. Contact me if you want a copy!
Evaluating LLM agents on finding shortest citation paths between academic papers.
Generates graphs for visualizing results and allows human player via web interface.
A tool to translate latex projects using LLMs, preserving structure and formatting.
Thesis project adapting the IDIR framework for 2D retinal image alignment using Implicit Neural Representations. Developed as my computer science degree thesis at UDC.
Written thesis and presentation materials available separately.
AI-powered CLI tool that merges data sources (.csv, .sql, .jsonl...) into a unified format. Automatically maps fields, infers missing headers, and rejects irrelevant information
Competed on Spain's biggest hackathon, where we created
VuelingConnect, a multiplatform app that creates a decentralized communications network using BLE, with a creative approach to ensure battery efficiency and avoid redundancies.
An open source and self-hosted alternative to
Offtop. Allows for uploading, tagging and sharing beats or packs with other artists.
Competed on 2 different projects during 3 days with a couple friends
-
cMatch, a web and chrome extension to find and managing clothes from Inditex.
-
skillhub, a hard and soft skills manager for organizations where users can post issues and guides.
A tool for benchmarking LLM agents by making them play
Escoba against each other. Using ELO rating system.
Using gym's occupancy data and maps, we built a simulation that allows managers to easily visualize it and identify problems
Image processing tool written in Rust, centered around
glitch and
corruption effects. Able to create custom filter pipelines or randomized ones, also added the ability to do image segmentation based on a text prompt based on
SAM2 and GroundingDINO
Placed top25, ok for the first time, learned lots about satellite data.
Simple Chrome extension for reducing sibilance in speech.
Trained a Lora for Llama 3.1. on a dataset of Miguel Anxo Bastos's conferences and articles.
Trained a model based on musicgen-small with drum breaks.
Performed an audiovisual live show at the event, using reactive and artifially generated visuals that morph with the music. Most of the music were my own productions, the visuals were made with StableDiffusion and TouchDesigner.
My personal dotfiles for configuring development environments and tools.