Most AI engagements ask you to commit to a programme before anyone can tell you whether it will work. These do the opposite: each one is a bounded piece of work, priced up front, that ends with a document you own and a decision you can actually make.
They are also not slide decks. Every one of them involves running something against your real data or your real system, and the findings come with the evidence attached.
Three phases, and you can stop after any of them
Phase 1 - Assess. The three fixed-price reviews on this page. Each answers one question about one system and ends in a report, a verdict and the machinery to check it again later.
Phase 2 - Build. Projects: custom models and systems, scientific R&D, research partnerships. Scoped from the assessment when there was one.
Phase 3 - Run. Handover, monitoring, retests on a regular cadence - or we keep operating the thing we built.
Two things worth saying plainly, because most consultancies leave them ambiguous. Buying Phase 1 commits you to nothing else: the report is written to be useful if you hand it to your own team or to another supplier, and we would rather you did that than bought a build you did not need. And a project does not require an assessment first - if you already know what you want built, skip to Phase 2.
Three reviews, one question: can this be trusted?
AI Readiness
Before you build. Which of your candidate uses of AI are actually feasible on the data you have, which are not, and what to do first. A scan that works from what you can show us, or a full assessment that runs baselines on your real data.
Model Validation & Deployment
After you build, against reality. Can you trust it, and how do you run it in production? Calibration, evaluation leakage, shortcut learning, slices and drift - plus what it takes to serve, monitor and pay for the thing once it is live.
AI Trust & Security
After you ship, against attackers. Red-teaming for LLM applications and agents - prompt injection, data leakage, tool misuse - plus a white-box review of the code around the model. Not a pentest: the failure modes that only exist once a model is in the loop.
Two of these look at a system you already run, and the difference between them is the failure mode. The security audit asks what happens when someone makes it fail; model validation asks whether it is quietly failing on its own. An LLM is still a model, so a chatbot can need both.
Each comes in two tiers. An entry tier - one system, the standard battery, a verdict - is €4,950 excl. VAT and typically takes two to four weeks. A full tier goes wide and deep, hands over the harness or the probe suite, and includes a retest: €24,500 excl. VAT, typically eight to twelve weeks. Both are fixed price against a written scope, and the number does not move because the work turned out to be interesting.
The clock starts when the data and the access are actually in place, not at signature, because that is the part nobody controls from our side. If it turns out there is not enough data to answer the question, that is a finding we write up - what is missing, how much would be enough, what can be said with what exists - not a project that quietly stalls.
What you can count on
Fixed scope and fixed price. You know what it costs before it starts, and the number does not move because the work turned out to be interesting.
Weeks, not years. None of these needs a long programme before it produces something worth having. Bigger work does exist - building the system, fixing what a review found - but that is quoted separately afterwards, never as a condition of getting started.
You keep the machinery. The probe suite, the calibration harness, the monitoring checks all go into your repository and run in your CI, so the next model gets the same scrutiny without hiring anyone.
- 1
Scoping call
Half an hour. What the system does, what decision depends on it, what a bad day looks like. This fixes the price.
- 2
Structured intake
Access, data, or documentation - as much as the question needs and no more.
- 3
Analysis
A standard battery for the offer, plus whatever your domain specifically demands.
- 4
Report and readout
Findings ranked by consequence, each with its evidence and a concrete remedy, presented to your team.
- 5
Handover
The tests come with you. A retest when you have acted on the findings.
Your data
Handing your data or your systems to an outside party to have them examined is a reasonable thing to be nervous about, so you choose the level at intake rather than discovering ours afterwards.
No AI on your data
No AI system sees your data or your code at any point. We still use AI to write generic code, the way we would with or without you - it simply never touches anything of yours.
EU-hosted AI
The default. Your raw data stays with you: the analysis runs on your infrastructure or in an isolated environment, and only structure and results come back. Anything that does go to a model runs in an EU region and is not used to train anyone's model.
Full access
Data and code through the same EU setup, under a data processing agreement, when you would rather trade that for speed.
Whichever level you pick, it is written into the engagement, along with where data lives, who can reach it, and when it is deleted. Client data never goes through a personal or consumer AI subscription.
An honest boundary
We are engineers, not counsel. These reviews produce technical evidence - measurements, reproductions, findings - which your legal and compliance people can use, and increasingly need. What we will not do is tell you that you are compliant with anything. That signature belongs to someone whose job it is.
What is never included, in any tier, at any price: compliance or legal sign-off, integration of anything into your systems, running or supporting the system afterwards, and fixing what a review found. The last three are real work we are happy to do - they are projects, quoted separately, because pretending they fit inside a fixed-price review is how fixed prices stop being fixed.
Where to go next
Model validation & deployment
For a scientific or industrial model already in use, or about to be.
AI trust & security
For anything with a language model in it.
Projects
When a review says build - custom models, systems, research partnerships.
How we build
The method underneath all of it - and what we do once a review says the system is worth building on.