Projects

A few side projects I’ve built to scratch my own itches, mostly around AI, personal knowledge management, and useful web tools.


personal-knowledge-base

Graph view of the personal knowledge base in Obsidian

A personal knowledge base that stops your AI context from evaporating between sessions. Based on Karpathy’s LLM Wiki pattern: instead of re-explaining who you are and what you’ve built every time you open a new chat, Claude Code reads and maintains a vault of markdown notes that compounds over time. Where RAG retrieves and forgets, a wiki accumulates and connects. Every session starts with everything you already learned.

Built with Obsidian, Claude Code, and markdown files versioned on GitHub.

View on GitHub →

Read the post →


tl-dr-bot

tl-dr-bot running in Telegram

A Telegram bot that acts as an intelligent filter for the links you accumulate during the day. Send any URL and it fetches the content, asks Claude whether it’s worth reading, summarises it in 3 bullets, classifies it by topic, and saves it directly to your Obsidian vault via the GitHub API. Includes deduplication. If you send a URL you’ve already saved, it tells you without consuming any API credits.

Built with Python, Claude (Anthropic), Jina AI Reader, and the Telegram Bot API.

View on GitHub →

Read the post →


second-brain-eval

Evaluation results comparing an LLM with and without a personal knowledge base

An experiment based on Karpathy’s second brain idea: does injecting your personal knowledge base into an LLM actually improve the quality of its responses? This project compares two conditions (LLM with personal wiki context vs. LLM baseline) and uses a third LLM instance as a blind judge scoring each response on precision, personal grounding, and practical utility.

Built with Python, Anthropic and OpenAI APIs.

View on GitHub →

Read the post →


minusculasmayusculas.com

The minusculasmayusculas.com text tools interface

A suite of free online text tools: case conversion, word counter, find & replace, and list sorting. No server, no registration, no tracking; everything runs client-side in the browser. Currently gets organic search traffic for Spanish-language text tool queries.

Visit the site →

View on GitHub →


totpdf.com

Free PDF tools that run in your browser. Merge, split, compress, organize, unlock and convert PDFs without uploading them to any server. In Catalan, Spanish and English, with no account, no ads and open-source code on GitHub.

Visit the site →


totimages.com

Free image tools that run in your browser. Convert HEIC, WebP and AVIF images to PNG or JPG, or turn them into PDF, without uploading anything to any server. In Catalan, Spanish and English, with no account and open-source code on GitHub.

Visit the site →