Olivier Abou. Teaching machines to talk since 1991.
Linguist by training, PhD in natural language processing (AI), author of more than 100 technical books and a tech journalist for 15 years. I've ridden every wave of the digital world (tech explainers, press, digital marketing, e-commerce) before focusing on AI tools applied to the real work of businesses.
What sets me apart: I come from the linguists' AI, the AI of rules and dictionaries. I apply it to today's models to build governed agents that follow your rules, rely on your sources and make far fewer mistakes.
Your rules first. The model second.
I started in AI back when the linguist was at its center: meaning was represented with rules and dictionaries. Today, models learn from billions of data points, and linguists have made way for mathematicians and engineers. I keep the rigor of the first school and the power of the second.
Dictionaries and rules
Knowledge written by hand, explicit reasoning. Predictable and verifiable, but rigid when faced with everyday language.
Billions of data points
Flexible understanding of language, fluent writing. Powerful, but capable of confidently making up an answer.
The best of both, in every agent.
I work a bit “old school”, with modern tools. I define the domain, the sources, the rules and the reasoning process. The language model only steps in where it excels: understanding and phrasing.
A narrow domain
Every agent has a defined mission: what it handles, what it declines, and who it hands off to.
Controlled knowledge
It draws on industry glossaries and knowledge bases validated with you, not on the whole internet.
Written, tested rules
What must always, or never, be done is defined in advance. Behavior is deterministic wherever it needs to be.
An enforced reasoning process
I set the steps: check the request, find the source, apply the rule, then answer.
The model, to understand and phrase
The large language model understands a freely worded question and writes a clear answer.
Verifiable answers, far fewer errors
Far fewer “hallucinations” (those confidently invented answers), a cited source, and an honest “I don't know” when the information is missing.
A hybrid, so-called neuro-symbolic approach: explicit rules and statistical models working together.
| Generic AI assistant | Libreo governed agent | |
|---|---|---|
| What it knows | Everything it has read, indiscriminately | Your sources, validated and scoped |
| The rules | Implicit, vary from one answer to the next | Written, tested, always applied |
| When it doesn't know | It may invent a plausible answer | It says so and hands off to a human |
| Calculations | Estimated by the model | Done by code, to the cent |
| Verification | Difficult | Every answer cites its source |
+Under the hood for technical teams
- Meaning representation
- Dictionaries, glossaries and relationships between business concepts, built with your team.
- Retrieval
- Retrieval-augmented generation (RAG) on a closed corpus: the agent only answers from authorized documents.
- Rules
- Coded conditions, deterministic calculations and checks; outputs are verified before they are sent.
- Reasoning
- An enforced sequence of steps, with intermediate checks and confidence thresholds.
- Models
- Large models via API, or open-source models installed on your premises, depending on confidentiality.
- Quality control
- Test question sets replayed with every change, human review of sensitive cases.
From words to machines, then to agents.
I didn't start with code, I started with words. That makes the difference when you design tools that have to understand, explain and hold a conversation.
- 1991–1995
PhD in natural language processing
Early work on a central question: how a machine can understand, reason and decide, using rules and dictionaries.
- 1996–2011
Tech journalism and radio
Fifteen years decoding tech news: the habit of explaining complex things simply.
- 1999–2016
100+ technical books for a general audience
Published in France by Microsoft Press, Micro Application, Eyrolles and others: Windows, Linux, web development, databases.
- Since 2011
Digital marketing and e-commerce
Acquisition, conversion, advertising, SEO: understanding online business from the inside.
- Since 2019
Libreo
Designing AI systems: data analysis, agents, automation, training.
- 2026
Custom software and AI agents in production
Training assistants, an automated digital agency, an AI setter, an accounting review agent, AI installed on client premises: systems running for their users. And the launch of libreo.io for US businesses.
Three convictions.
The most useful AI is the kind you don't notice: it fits into everyday work, respects the way you do things and gives you time back.
Understand the business before the tech
I start from your workflows, your documents and your constraints. Technology comes second.
Explain every choice
You know what the tool does, why it does it and how to evolve it. No black box.
Measure what it delivers
No tokenmaxxing: a tool is judged by the time it saves and the results it produces, not by how much AI it consumes.
AI and automation expert at the 3Xpansion collective
Within 3Xpansion, a French growth agency for companies with more than €500K in annual revenue, I work alongside experts in business strategy, marketing and customer acquisition.
Let's talk about your project.
A no-commitment call to see whether I'm the right person to help you.