Fractional and freelance AI specialists
A register you can check.
You cannot tell from a profile whether an AI engineer is any good, and neither can a marketplace. So every specialist here is assessed against a rubric we publish, on evidence we can point at, and the profile says plainly what has not been verified.
The register is being assembled now. Both routes go to a real person, not a waiting list.
Worked example, not a real specialist. Generated by the live scoring engine on this page load, not edited by hand. The register has no members yet, so there is nobody real to show you. The row reading not verified is the point.
Why a specialist register, and not a marketplace
A general marketplace solves discovery. It does not solve the problem you actually have, which is that you cannot tell from a profile whether an AI engineer is any good, and neither can the platform. Star ratings measure whether past clients were happy, which is not the same thing.
So we do the part that is hard. Every specialist is assessed against a rubric we publish, on evidence we can point at. When something has not been verified, the profile says so rather than leaving you to assume. An unarguable badge is worth nothing.
Read the rubric in full, including the four things vetting deliberately does not do.
What the name means
Two days a week>an evaluation pipeline your team can run without you
Six weeks of a governance lead>an audit you can actually pass
Both sides of an inequality must be true, specific and attributable, or we do not use the device. These two describe the shape of the work, not a result we are promising.
How it works
Tell us the shape of the work
The discipline, the seniority, whether it is a day a week or a three month push. A job description works too: paste it and we will structure it for you.
Get a shortlist with the reasoning attached
Not a match percentage. Each name comes with what was checked, what the evidence was, and what is still unknown. You can disagree with any of it.
We introduce, you engage directly
The contract and the money stay between you and them. We are not in the middle of it, and we never take a cut from the specialist.
The disciplines
Fourteen areas, chosen because each one is genuinely hard to judge from a CV. That difficulty is the reason this exists.
AI strategy
Decides what a business should and should not build with AI, and in what order.
Everyone can talk about strategy. The tell is whether they have killed a project that was already funded.
AI engineering
Builds the systems that put models into production and keeps them there.
A demo proves nothing. What matters is what they did when it failed under real load.
LLM applications
Designs and ships products where a language model is the core of the product.
The gap between a working prototype and something that survives real users is where most of the work is, and it does not show in a portfolio.
Agents and retrieval
Builds agentic workflows and retrieval systems that return the right thing.
Retrieval quality is measurable, but almost nobody measures it. Ask for the numbers and most cannot produce any.
MLOps and platform
Owns the pipelines, deployment, monitoring and cost of models in production.
The work is invisible when it is done well, so the evidence is in incidents and spend, not in features.
Data engineering for AI
Gets the data into a shape a model can actually use, reliably and repeatedly.
Most AI projects fail here rather than at the model, and the people who know that are rarely the loudest in the room.
AI governance
Builds the policies, records and controls that let an organisation show how its AI is run.
A framework can be recited. Having actually taken an organisation through an audit is a different thing entirely.
AI risk and compliance
Maps obligations under regimes such as the EU AI Act and ISO 42001 onto what a business is really doing.
The field is young, the certifications are new, and a confident summary of a regulation is not the same as having applied it.
Safety and evaluation
Designs the tests that say whether a system is good enough to ship.
Writing an eval is easy. Writing one that would actually have caught the failure is not.
Red teaming
Attacks a system deliberately to find how it breaks before someone else does.
Claims are easy and evidence is often confidential, which makes references and method the only real signal.
AI product management
Decides what to build, for whom, and how to tell whether it worked.
The discipline is barely five years old, so titles say very little about what someone has actually shipped.
Fine-tuning and adaptation
Adapts existing models to a specific domain, and knows when not to.
Knowing when fine-tuning is the wrong answer is worth more than knowing how to do it, and it is harder to demonstrate.
Training and enablement
Gets a workforce genuinely capable with AI tools, rather than merely trained.
Attendance is easy to report. Changed behaviour six months later is the thing that matters.
Applied research
Works on problems where the method is not settled yet.
Publication record is a signal but not the signal, because the applied question is whether it survives contact with a product.
Who pays, plainly
Hirers pay, on introduction or by retained search. Specialists are never charged a fee, for anything. That is not generosity: charging a work-seeker to find them work is prohibited under UK recruitment rules, and any platform that does it is one you should look at twice.
We also never hold or move money between you and the person you hire, and we do not determine employment status. You contract directly. That keeps the arrangement simple and keeps us out of places we have no business being.
Try the thinking before you trust the register
Assay is the instrument behind the vetting, and a free version is on this site. Paste a job post and it gives you a straight read: what fits, what does not, and what it cannot tell from the information available. No account, and nothing you paste leaves your browser.
Open Assay