I'm obsessed with the synthetic personality of AI models and share it at IA para gente curiosa, an academic, practical and didactic space.
I build what I need and publish it: npm packages and PyPI, apps for Linux and macOS, the odd iOS app, SaaS, APIs and a good pile of projects and POCs.
Almost everything revolves around software engineering: systems, agents and tools like Alfred Dev, PSPO Agent, Claude Island or Crupier. Also LegalTech (regulation, VAT and privacy), offensive and defensive security and social network analysis. I also contribute to Open Source (with patches in Hermes Agent and Hermes Studio, among others), explore mind-blowing conjectures and write about technology.
Agent harness engineering under one thesis: the models are not the variable, the harness is. Design of the deterministic layer that validates, authorizes, executes and logs every action the model proposes.
Claude Code plugin that turns an isolated session into a governed flow: 19 agents, 25 commands, 62 skills and per-project SQLite memory.
A pack that turns a Hermes install into a complete scientific operation: PRISMA systematic reviews via Telegram or CLI, with state outside the chat and an editorial package.
A QA-agent harness over repositories: it coordinates quality, security and CI/CD skills on GitHub, with static analysis, reports and actionable issues.
Product Owner plugin for Claude Code: Given/When/Then user stories, sprints with an AI factor and publishing to Trello, Notion or GitHub Projects.
Prepares classes from your real syllabus: lesson plan, slides, exercises, wrap-up and assessment, with citations to the PDF page.
Native macOS control surface for Claude Code, attached to the MacBook notch: sessions, permissions and subagents at a glance.
Linux top-bar indicator watching your AI coding agents (Kimi Code, Claude Code, Codex): "needs you" alerts with the exact question, live tokens, context % and daily cost.
Turns any public GitHub repo into an exhaustive brief to rebuild it with coding agents: 8 specialist agents working live.
Multi-provider LLM orchestration in Python: the LLM proposes the plan and policy decides whether it may run.
Self-hosted Zero Trust layer for MCP servers: hides tools, blocks calls, requires human approval and audits every decision.
Anonymization of Spanish legal documents: a deterministic engine plus an optional AI layer, validated on more than 100,000 real documents.
Reads your legal brief, extracts every regulatory reference and checks against the BOE and EUR-Lex whether it is still in force, was amended or has been repealed.
Classifies batches of invoices, prepares the VAT ledger and reconciles the Spanish Modelo 303 with a local, traceable and reviewable flow, with AEAT traceability.
Which court has jurisdiction? Determines the competent Spanish court by postal code, legal matter and amount, with LexNET code and legal basis.
API to query regional, national and EU legislation in natural language with AI, citing article by article from official sources (BOE and EUR-Lex).
Spanish legal acronyms, expanded and audited: 647 verified, with BOE and EUR-Lex resolution.
Automated technical audit of cookies before consent, with web and PDF evidence reports for GDPR.
An Odoo 19 module that turns supplier invoice attachments into reviewable drafts: extracts vendor, lines and amounts with AI and never posts on its own.
API of pharmacy openings, closures and relocations in Spain, with daily ingestion from 19 official sources.
A team of 6 LLM agents that diagnoses mainframe batch errors in COBOL→Java migrations for a Spanish Tier 1 bank: upload a ZIP with logs and stacktraces and it returns root cause and correction plan.
Watermarks for PDFs and images 100% in the browser: your files never leave your device.
Camouflages running processes by renaming them with system service names, invisible to standard process scanning. Linux and Windows.
Camouflages VirtualBox virtual machines so they are not detected by systeminformation. Changes DMI/SMBIOS, CPU, disk, MAC and PCI devices with real hardware profiles.
Working PoC of CVE-2026-64638 (XSS2Shell): a pre-auth XSS-to-RCE chain in WordPress Core, with a Docker lab and a Python attacker server.
Open facsimile to understand, build and defend AI systems: 12 installments, 113 chapters and a 1105-term glossary.
Academic map of 385 papers on synthetic personality in LLMs, critically reviewing method, results and limits.
Interactive web presentation of the AI workshop for curious people: 438 navigable slides, from foundations to agents in production.
OSINT observatory of message propagation on X: communities, emitter nodes and bridges, with open data.
Open REST proxy for TUSSAM's internal API: Seville's 967 bus stops in JSON, with caching and resilience.
Linux top-bar indicator watching your AI coding agents (Kimi Code, Claude Code, Codex): "needs you" alerts with the exact question, live tokens, context % and daily cost.
Order your GitHub profile repositories the way you want: up to 100 pins, beyond the native 6.
Drop-in replacement for minimatch: up to 17x faster powered by picomatch, passes the 402 original tests and eliminates CVE-2022-3517.
Browsable catalog of 99 technical roles × 9 career levels: 891 bilingual EN/ES definitions, straight from the PyPI package.
Numerology attributes mystical significance to certain numbers based on apparently deep patterns. Rather than refuting such claims case by case, this work adopts the reverse strategy: openly constructing a rigorous mathematical system in which every natural number converges to 7 by design, and developing its complete theory with the same apparatus a conventional paper would display. Universal convergence, an exact closed form, optimal bounds, extremal values, uniform distribution over dyadic octaves and exhaustive verification up to 10⁷: a demarcation experiment between mathematics and pseudoscience.
20 min
A University department posed a concrete problem to me: moving from model-assisted exploration to a PRISMA review that is methodologically defensible, technically auditable, and deliverable as a real editorial package. That's how I built Hermes: an agent system with rules, files, commands, and material controls. DOI-first, mandatory full text, editorial gate, and memory outside the chat.
14 min
Added thinking content blocks support to the Anthropic Messages API in llama.cpp. Models like DeepSeek-R1 and Qwen3-Thinking now work correctly with Claude Code and other clients following the Anthropic specification.
5 min
Running