Understanding it. Engineering it. Exploring what comes next.

AI is changing how we think, build and interact with the world.

01 · AI & Business

From AI potential to business value.

AI is worth what it changes in a business, not what it can do in a demo. The hard part is rarely the model; it is the decision around it.

Founders, architects, technical leaders and decision-makers meet the same few questions before any choice of technology: where it pays, what it costs to run, what can go wrong, and whether to build, buy or wait.

01Where it paysStart from a process with a cost you can measure, not from a model in search of a use.
02What it costs to runPrice the answer, not the licence: volume, latency, and the people who check the output.
03What can go wrongWrong answers, leaked data, quiet drift. Each needs an owner before launch, not after.
04Build, buy or waitBuy the commodity, build what sets you apart, and wait where the ground is still moving.

Article · 14 November 2026 · 20 min readAI-First Is Not the Same as Outcome-FirstWhy the sequence of AI adoption matters more than the technology itself.Read the article

In the works · ProjectThree levels of CX: automation, AI and agentsComing soon

02 · AI Engineering

Beyond the model.

A model is the smallest part of an AI system. What decides whether it holds up in production is everything around it: the data it reads, the context it is given, how it is served and orchestrated, what its agents and tools may do, and how it is measured and governed.

These are not six steps in a row. A request passes down through four layers and its answer comes back up; two concerns run across all of them, and what evaluation finds flows back into the data, the retrieval and the prompts.

06Governance & SecurityAccess, redaction, logs, human review and policy, enforced at every layer.

04Serving & OrchestrationA request enters hereRoutes each request, calls models, tools and agents in order, holds the latency and cost of an answer, falls back when a step fails.Trade-off · Every hop adds latency; every fallback is one more path to test.
03ModelsHosted or your own, chosen per task rather than once for everything.Trade-off · Capability against cost, speed and control of the data.
02Retrieval & ContextSearch, ranking, and the context an answer is built on.Trade-off · More context is not better context: precision beats volume.
01DataSources, lineage, freshness, and who may read what.Trade-off · An answer is never fresher, or cleaner, than what it was built from.

05Evaluation & ObservabilityFixed cases before release, traces and scores after it, at every layer.

Governance & Security, beside the stack: access, redaction, logs, human review and policy, enforced at every layer.

Evaluation & Observability, on the other side: fixed cases before release, traces and scores after it, at every layer.

Four principles

Data before models.Most AI failures are data failures. Lineage, freshness and access come before any choice of model.
Evaluation as a test suite.Every change to a prompt, a model or an index runs against a fixed set of cases before it ships.
Answers you can trace.Retrieval that cites its sources, so a wrong answer can be followed back to where it came from.
Guardrails at the boundary.Limits, redaction and review sit outside the model, where they can be tested like any other code.

Article · 28 November 2026 · 21 min readExclusive until 28 November 2026Anatomy of a Production AI SystemFrom models and prompts to reliable, scalable systems: the layers around the model, and what changes when real users, real costs and real risks arrive.Exclusive, on request

Guide · 24 pagesExclusive until 28 November 2026The Anatomy of Production AIFrom models and prompts to reliable, scalable systems: the eight layers around the model, the three trust boundaries between them, the question to ask at each, one risky case run through them all, and a checklist for before real users arrive.

03 · Frontier

What comes next?

Agents that act on our behalf, robots that share our spaces, machines we speak to instead of type at, interfaces that disappear into the places we use them. Some of it is already working somewhere; most of it is still a question.

Frontier is where I explore those possibilities, with curiosity and with the questions I ask of any system: what does it need, what does it change, and what could go wrong? What is built is marked as built; a concept is marked as a concept.

Concept · ProjectFuture Edge StationAn immersive experience, built like a small game: you travel to a universe a little ahead of ours and walk through it, meeting its technology where it would actually be used.Open the project

In the works · EssayAI can see through wallsComing soon

04 · Podcasts & Workshops

Ideas are better when challenged.

Ideas tested in the room: workshops and talks for the people who decide, and a podcast for the questions worth arguing about, with people who see them differently.

Mouad Krikbou at one of the workshops, pointing to a slide of nested circles: AI, machine learning, deep learning, generative AI and large language models.

Workshops · Vodafone · 2024

Introducing AI to C-suite executives

In 2024, as AI was booming, we introduced AI to C-suite executives in workshops in Munich, Hamburg and Stuttgart.

Recaps, in GermanMunich (the organiser's recap, opens in a new tab)Hamburg (the organiser's recap, opens in a new tab)Stuttgart (the organiser's recap, opens in a new tab)

Podcast · In preparation

Short conversations on AI, engineering and business, under three minutes each.

Suggest a topic or a guest

The work behind it

Selected AI work

Essay · 15 September 2025 · 8 min readThe truth and nothing but the truthThere is no absolute truth accessible through AI. It falls upon us to connect the dots, critically evaluate information, and discern whether it represents a fragment of the broader truth.Read the essay

An idea worth building?

Or a decision on the table, or a question worth arguing about. I'd be glad to hear it.

Get in touch