Two decades designing systems, setting technology strategy and helping organisations modernise at scale, while still hands-on in the architecture and the codebase.
Madrid, Spain · open to international roles and remote engagements · Spanish, English, Italian
CTO with two decades in architecture and engineering leadership across enterprise, consulting and startups. Most recently built AI products end to end as CTO of an AI-driven energy business. Before that, ran the global architecture function at SGS, reporting to the Group CIO, as design authority for cloud, modernisation and engineering standards across an IT organisation of more than 800 projects. Works equally at executive level and in the codebase, and is as comfortable running a large engineering organisation as delivering solo.
Earlier career, 2001 to 2015. Solutions architect and team lead at Siemens (MDM and BPM platforms for the West Europe cluster), PwC (.NET architecture and SDLC) and VASS (BT, Repsol, Isban), plus contract architecture for Banco Popular and Banco de España. Started as a developer in 2001; freelance English and Spanish technical translator from 1996 to 2002.
SIGINTENT: architecture reviews, AI strategy sprints and fractional CTO work for founders and executive teams.
Public repositories, each with CI, tests and a README that says what was measured. Everything below runs locally first and treats a model's output as a claim to verify, not a fact.
Analyses a document corpus and the infrastructure constraints, then recommends a complete RAG configuration: embeddings, chunking, hybrid retrieval, vector store, reranker, sizing, cost and latency. Every number carries its provenance, and --validate measures the recommendation against ground-truth queries before anything is indexed.
A boardroom of AI agents declared in YAML: hostile round, chair-led deliberation, coaching and synthesis, with a refine loop that sends the revised idea back to the board. Seats can be grounded in a local library or a company's records through MCP, and every verdict carries confidence, caveats and open questions.
A programme for an in-house AI assurance capability: bias, explainability, adversarial robustness, evaluation frameworks, differential privacy, the EU AI Act timeline, agentic safety. Prose under CC BY, code under MIT, and a companion lab that runs the evaluations against a model you choose.
The step every knowledge-base project skips: point it at a folder and it reports what the documents are worth as text. Per-file metrics, OCR quality, PII with checksum validation, Markdown-convertibility verdicts, duplicates and versions, topic clusters, and a curation plan with one recommended action per file.
A complete implementation of OpenAI's tokenizer in C, written as teaching material. Differential tests against the Python package (3,000 comparisons per push) and fuzzing of both parsers found real bugs, and the README says which. The eight-chapter tutorial became the book below.
An opinionated portfolio tracker that makes you write down when to quit: abandonment criteria and pre-mortems up front, subjective completion against the figure computed from tasks. Driven by an agent through its MCP server rather than by a chat window, with a Claude Code skill that keeps each repo's record honest.
One pure core library exposed four ways: about 130 HTTP routes, a CLI, gRPC and three NuGet packages. PDF, DOCX, XLSX, PPTX, CSV, EPUB, MOBI, email and more, with PII detection and hashing. The AGPL PDF engine is isolated in its own package so the core stays MIT, and the pipeline produces an SBOM, runs CodeQL and dependency review.
Perceptual image hashing bit-compatible with the Python imagehash library, verified against a golden dataset. A pixel budget against decompression bombs, first-frame-only decoding and strict parsing, with a written threat model. Small on purpose, and finished.
A 2014 library rewritten from scratch in 2026 under its original NuGet id, keeping only what the base class library still lacks: string metrics (Levenshtein, Damerau, Jaro-Winkler), sequence operators and functional helpers. Every README example is asserted by a test.
A deterministic generator of search-operator queries for OSINT and defensive exposure discovery. Two engines, one in JavaScript and one in Python, kept identical by 153 shared golden fixtures. No model involved, by design: when the output must be exact, a rule beats a guess.
The companion editor to the XML-Structured Prompting book: versioned prompts with diff, templates, labels, a command palette, projects with membership and audit, SQLite storage and a CI with coverage floors that only move up.
Tooling for agents, published: an SEO and generative-engine audit with SARIF output, a BPMN mnemonic, a repository badge detector, an llms.txt generator, a break timer built on hooks, a vibe-to-functional rewriter and a commit poet. Each one solves a problem the author actually had.
What a BPE tokenizer does and how to build one that matches the reference bit for bit, with the tests that prove it.
A method for prompts that survive handover: explicit structure, constraints and evidence, for teams that ship with language models.
In English, with Spanish adaptations on LinkedIn. More in the pipeline on irreversibility, agentic systems as distributed systems, and what a RAG advisor teaches about evaluation.
Eight essays on what the classical tradition (Seneca, Marcus Aurelius, Pascal, Ortega, Byung-Chul Han) has to say about living and building with machines that think.
A degree in English philology and six years as a technical translator came before the first line of production code in 2001. Since then: developer, architect, head of a global architecture function, CTO, with detours through a master's in data engineering, an INSEAD programme on business strategy and a digital MBA for technology leaders. On the side, philosophy, music, photography and languages. The pattern is not breadth for its own sake. It is what the industry now calls an M-shaped profile: deep in architecture and AI, and able to move between product, strategy and organisational design without losing the thread.
Most technology decisions get answered before anyone has agreed what the question is. That is how a defensible architecture becomes an expensive one.
The hub: one screen that routes to the writing, the code, the books and the advisory. Start here if you want the whole picture.
aelena.comThe advisory practice: fractional CTO, architecture review, AI strategy sprint, advisory retainer. Method and essays, in English and Spanish. Start here if you have a decision you cannot unmake.
sig-intent.comantonioelena@gmail.com+34 686 398 711