ls ~/research/ai
Artificial Intelligence
An engineering service centered on the harness: the deterministic layer that wraps the model and validates, authorizes, executes and logs every action it proposes. It addresses five use cases —taking a demo to production, adding traceability to what already runs, passing an audit, building a harness from scratch and assessing whether one is actually needed— with a one-to-two-week initial assessment that may conclude none is required. The design works on tool contracts, context control, verification before acting, state management between steps, cost and policy limits, concurrency control and full execution lineage to reproduce any decision. Direct contact with the engineer over encrypted email, no CRM or sales sequences.
Standalone Claude Code plugin orchestrating 19 real agents (10 core + 9 optional) with 62 skills across 15 domains. Includes 6 quality-gated flows (feature, fix, spike, ship, audit, quick), 26 namespaced commands + /alfred, persistent memory in local SQLite (WAL + FTS5), real-time dashboard, operational PM layer (SonIA), design system (Selina) and external technical second opinion (Lucius). Quality gates force you to prove the work with verifiable evidence before advancing. No external dependencies or remote services: everything runs locally on Python 3.10+, Node.js and SQLite.
A reproducible runtime (Docker, clean-room validated) that adds the complete systematic-review flow to Hermes in twelve phases with a single material state: wizard intake, question decomposition, protocol before corpus, logged multi-source search (OpenAlex, Crossref, Semantic Scholar, arXiv), DOI-first with deduplication, OK/KO screening with justification, full text with hash and reconstructed reading (Poppler + Docling + OCR), deep field-aware extraction (biomedical, technical, social sciences, education, management), synthesis, independent review and a deterministic publication gate. Ships a compiled PDF manuscript, editable LaTeX, modular MD, CSVs, figures and a ZIP, with a watchdog that resumes phases after restarts and material memory in the workspace and Obsidian.
Product Owner plugin for Claude Code that professionalizes AI-assisted PO work. Analyzes requirements, writes user stories with Given/When/Then format and acceptance criteria, plans sprints incorporating an AI complexity factor and publishes the result in Trello. Designed for teams that want assisted PO without losing Scrum method.
Critical literature review on synthetic personality in language models: 385 unique papers with full entries (method, results, limits and what each study does not prove), organized into 6 research lines: psychometric validity, trait induction, agents and identity, collective behavior, applications and bias, and governance. Published at syntheticpersonality.com.
Multi-agent system working sequentially: Explorer (analyzes structure and classifies tables), Observer (monitors metrics and detects anomalies), Architect (proposes indexes with impact estimation), Gardener (detects bloat and schedules maintenance), Partitioner (recommends strategies for large tables). 4 autonomy levels depending on environment. Does not modify database, only analyzes and proposes.
Automated batch error analysis platform for mainframe environments running JPLI applications (COBOL to Java migrations), built for a Spanish Tier 1 bank. The user uploads a ZIP with the full evidence of a batch error —organized in nested directories per run— and a dispatcher walks that hierarchy, classifying each file by naming convention (JCL scripts, stdout logs, XML stacktraces and the JPLI/Java source tree extracted from the stacktrace itself) and routing it to the matching agent. A team of 6 specialized LLM agents works in parallel: they analyze the JCL scripts, the stdout logs, the exception hierarchy in the XML stacktraces and the JPLI/Java code involved; a fifth agent correlates every source, rebuilds the error timeline and determines the root cause with an associated confidence level; and a sixth generates the final technical report —executive summary, root cause, supporting evidence, impact and an actionable correction plan— following the structure agreed with the client. The whole flow is traced with centralized, level-based logging. Confidential client project.
- aiEasyTXT
TXT-first macOS editor with integrated AI assistants (OpenAI/Claude): correction, idea expansion and suggestions. Markdown/Mermaid with split view.
Outreach site that translates my PhD research on synthetic personality in AI into accessible language for curious people. Explores how the personality of language models is formed, measured and biased, with practical examples and no academic jargon. A bridge between the academic research in LLM Synthetic Personality Review and the general public.
Hands-on notebooks from the «Inteligencia artificial para gente curiosa» facsimile, organized by installment and runnable on Google Colab. Each notebook explains one concept step by step with code and real outputs: a hand-built neural network with NumPy, overfitting live, classic ML, BFS/DFS/A*, Sudoku as a CSP, planning (blocks world), minimax/alpha-beta, Q/K/V attention by hand, temperature/top-k/top-p, quantization, a mini-GPT from scratch, tokens and cost, structured outputs and a minimal ReAct agent.
A complete workshop as a keyboard-navigable web slide deck, updated May 2026: foundations (neuron, backpropagation, tokens, embeddings, quantization), classic AI (BFS/DFS/A*, SAT/CSP, PDDL, minimax/MCTS, ontologies/SPARQL), architectures (Transformers inside-out, MoE, SSM/Mamba), RAG and multimodal RAG, fine-tuning, agents in production, MCP, evals, computer-use, red-team tooling, LLMOps, agentic security and vibe coding.
An open skill for AI agents that turns a folder or ZIP of invoices (PDF, photos, Facturae, XML) into an orderly, local and reviewable VAT classification process: it inventories without touching the originals (type, size, SHA-256), extracts text and runs resumable OCR when needed, separates file, invoice, operation and row, and classifies through a traceable AEAT decision path that requires reaching a terminal case. It only asks what is decisive (country, establishment, exemption, reverse charge) and, after the tax profile is reviewed, exports separately an audit ledger and an XLSX matching the AEAT 2026 record-book structure to reconcile the Modelo 303. An independent, unofficial product; it does not file returns nor replace official validation.
An independent module for Odoo 19 Community (LGPL-3.0) that processes one or several attachments per email, separates invoices from signatures, logos and duplicates, and extracts vendor, NIF/VAT, dates, line items, amounts, taxes, currency and IBAN. Works with OpenAI, OpenRouter, local or Cloud Ollama and OpenAI-compatible servers, with a primary and a fallback model. It only creates supplier invoices as drafts —it never posts entries, orders payments or creates new vendors—, requires an exact match with an existing vendor and keeps per-document history of attempts, states and alerts for the accounting team to review.
- aiAgente Q
A team of QA agents that works alongside your development team: it reviews GitHub repositories and aligns with your continuous integration systems. It coordinates domain-specific quality and security skills, runs static analysis and delivers portal reports and actionable GitHub issues, without launching the product. Designed to plug into the development flow with traceable evidence. Access by invitation.