Applied AI leadership, built on how people actually learn.
I lead technical documentation and put AI to work in it: an agent chosen by evaluation, traceable change methods, quality governance and a career path that treats AI as a skill. Some of this work is in production and some is still a proof of concept or a proposal; each case states its status and what it has not proven yet.
I wanted an agent that answers from our versioned technical manuals and always cites its source. I had three architectures built and evaluated, and chose with data, not opinion.
4.46 → 4.85
LLM-judge score (1–5), v1 RAG against v2 full context, on 26 golden questions
12/17 → 15/17
answerable questions with a grounded answer, v1 against v2
97%
Recall@10 for v3 on 34 paraphrased questions in Spanish and English
04 decisions of my own · Proof of concept · not in production
Change requests reached us as emails, tickets and version diffs thousands of lines long. I turned them into a versioned, traceable method and applied it to three product releases.
115
changes traced across 3 releases (43 + 56 + 16)
54
questions for subject-matter experts
27 · 14
Jira stories traced and reviewer comments incorporated
04 decisions of my own · In production · 3 releases
My team started using AI to draft, translate and review, and nobody logged the errors it introduced. I set up checkpoints, an error log and two roles so those errors change the prompts.
3
checkpoints, one per stage: draft, translation, pre-export review
5
workbook tabs: instructions, three stage checklists and an error log
2
new roles: TW Reviewer and TW Prompt Owner
04 decisions of my own · In production · the area's only formal AI-quality record
HR's 360° evaluation lived in a 10-sheet spreadsheet. I directed three architecture generations and an analysis that cut one questionnaire from 62 to 26 questions and corrected how people are scored.
62 → 26
questions for one role once the role-to-level matrix was recovered (−58%)
149 → 57
workflow nodes after the move from the first to the second architecture
41 · 7
findings fixed in two adversarial verification rounds, 7 of them blocking
04 decisions of my own · In production · Gen 3, self-contained
My documentation team had no explicit career path. I designed a four-level framework, now approved by HR, in which working well with AI is assessed like any other skill.
4
levels, with 27, 27, 31 and 34 skills
18
functions of the role, with at most 12 active per level
2
AI competencies, one function and one hard skill
05 decisions of my own · Approved · HR is implementing it
Smaller in scope, in production or in review, each with what it has not measured yet.
Status: In production
A documentation feedback system
Support, Delivery and customer success report documentation errors through one form, and a monthly report closes the loop after approval.
10 · 36
products and documents in the catalog, in English and Spanish
9
KPIs defined, from volume by type to backlog older than 30 days
< 1 min
to fill in a report
Assisted: Claude Code built the static form and documented each flow step by step, under my rules.
Status: In review with Delivery and Support
One manual per situation, not three per format
I directed a cut of about 65% to a legacy console manual without losing procedural content. For the new web console, I replaced three format-based manuals with one organized by the reader's situation.
~65%
shorter legacy manual, against a target of 30–50%
3 → 1
manuals, from one per format to one per situation
118 · 0
internal links in the web manual, none broken
Assisted: Claude did the text reduction, the cross-check against sources and the Word file, index and links under those rules.
How I work
Five rules I hold myself to.
Each one comes from decisions I made in real work between July and October 2026.
How the work is split
I decide and direct; Claude executes under my rules.
01
Ask before executing
I ask before I act, and I made it a rule of my own. In a documentation audit I escalated 7 decisions to the PM instead of settling them by assumption. In my change-plan method, a contradiction between sources is logged and escalated, not resolved on my own judgment.
02
Measure abstention, not only accuracy
A system that answers confidently when it should not is the riskier failure. My documentation agent's golden set included questions with no answer in the documentation, about another product and about another version. Both versions handled the trick questions 9 of 9.
03
Think about the cost of reversal
I weigh a decision by what it would cost to undo later. In the 360° portal, adopting the corrected model early costs almost nothing; changing it once the whole organization has been evaluated means telling everyone the model changed.
04
Protect data by design, not by omission
I ruled out Pandoc for converting 194 manuals, because our privacy policy does not allow it. The converter Claude Code built for me uses only Python's standard library. The rule was simple: content does not leave the team.
05
Stop the automation when you lose control
In the documentation agent I chose a fixed flow over a tool-using agent, because the agent could not be forced to call its tool. Automation is worth keeping only while I can still verify what it does.
Lab
Proofs of concept, proposals and prototypes.
Two groups: what I have directed or proposed at work, then prototypes I design on my own.
01 · At work
At work: proofs of concept and proposals.
Work inside the company that is not in production. Each one says whether it is a proof of concept or a proposal.
Proof of conceptProcess proposal not implemented
A request system with service levels
A Kanban board for documentation requests, with SLIs, SLOs in three scenarios and SLAs per internal client.
What it includes
A request view for PMs and a five-state Kanban for writers
4 SLIs and P90 SLOs in three scenarios: conservative, standard and ambitious
SLAs per internal client, from PMs to support priorities
Assisted: Claude Code built the board under my rules.
Rebuilding an operations manual, with one house style
Product C's operations manual had nearly twice the words of a comparable manual and read like another company's. I proposed splitting it by audience and adopting one house style across products.
What it includes
A style comparison between two products' manuals: headings, procedure labels, how the reader is addressed and note density
Volume measured by chapter, with four redundancy patterns diagnosed
A two-document plan, Configuration and Operation, sharing one single-source preamble
Assisted: Claude measured the volume, diagnosed the redundancy and drafted the tables of contents.
ProposalDelivered · decisions under PM review
An audit of published documentation
Nobody knew which version of each document was official, or how far behind the published PDFs were. I had three repositories mapped and escalated seven decisions to the PM instead of guessing.
What it includes
A map of the documentation for Product C's add-ons across 3 repositories: work in progress, team files and published PDFs
Word sources compared with published PDFs: 5 of 5 outdated, the worst by about 8 months
A table for the PM with 7 decisions escalated, each with its context and question
Assisted: Claude inventoried the repositories, compared the versions and drafted the emails.
ProposalSubmitted to HR
Defining the documentation role and how it is evaluated
HR asked me to describe the documentation leadership role and how it should be evaluated. I answered with ranked responsibilities and an evaluation design built on four fairness conditions.
What it includes
A role description with 6 final accountabilities
Responsibilities ranked by strategic importance, not by time spent
Mirrored indicators of success and difficulty
An evaluation design with 4 fairness conditions
Assisted: Claude drafted the text in English and then in Spanish, with my corrections.
02 · Personal lab
Prototypes, not products.
Six generic prototypes, unrelated to any employer. I designed and orchestrated each one to test an idea before it reaches a real team.
Prototype
Micro-learning Content Factory
Topic in, broadcast-ready video script out. I designed the pipeline so the instructional-design expertise lives inside it.
A script schema that enforces Gagné's Nine Events of Instruction
An intake web form where each field explains why it matters pedagogically
What I learned
Specialized sub-agents beat one mega-prompt. And making a private automation public and real-time (webhooks instead of forms, a database instead of an inbox) is its own discipline.
Prototype
Feedback Architect
Every learner comment is read on arrival and returned as sentiment, technical gaps and risk level.
A form with immediate persistence
An n8n "Silent Analyst": a single-purpose Gemini call
What I learned
A single-purpose judge prompt that returns one schema is more reliable at scale than a general summarizer that is also asked to classify. I design for what the admin needs to see now, not for what data we happen to have.
Prototype
CyberCoach, an adaptive assessor
A security assessment that tests readiness through live scenarios, then acts on the result.
Agent A: profiles the learner's role, generates situational scenarios live and scores competencies per topic
Agent B: when the conversation closes, writes a personalized follow-up email and curates resources for the gaps
What I learned
Splitting "assess" and "act on the assessment" into two agents simplified both prompts. The hard part was sequencing two asynchronous agents around one session ID.
Prototype
Post-Learning Intelligence
One training dataset, four stakeholder-specific reports, generated in a single run.
ROI calculated from KPI deltas
Four reports in one run: executive, ROI and finance, pedagogical audit, and operations alert
What I learned
Forcing reliable section boundaries in one LLM response took more than plain delimiters, which got swallowed. A marker that is not natural language solved it.
Prototype
Talent Mobility Orchestrator
A two-way conversational copilot: it fills a role from the workforce, or finds growth paths for one person.
An n8n AI agent that calls data tools on demand: profiles, DISC and open positions
Gemini scoring four dimensions: transferable skills, behavioral fit, performance trajectory and development investment
What I learned
When an agent needs precise data, build the precision into the data layer (RPC), not into the prompt.
Prototype
Orbit Docs, a RAG system
Answers from long product docs, with the source document and the exact passage on screen. Questions outside the docs get "not covered".
Gemini embeddings and a Pinecone index
An n8n agent that uses Pinecone as a tool
What I learned
Citation discipline mattered more than retrieval quality. An agent that finds the right passage but does not show which one is no more trustworthy than one that hallucinates.
About
An applied-AI leader, trained in how people learn.
I am Senior Manager, Technical Documentation in enterprise cybersecurity (digital fraud prevention). I own documentation for a line of fraud-prevention products and lead a team of two technical writers. My master's is in neuroeducation, so I start every AI system from how people learn and decide.
Degree
Máster en Neuropsicología, Inteligencias Múltiples y Mindfulness
IMF Smart Education · 2021–2024
Master's degree in neuroeducation · 60 ECTS · 1,500 h · GPA 9.25
