The AI you bought
is not the AI
you were sold.

Every AI product ships in two versions. There is the one in the documentation, calm and precise. And there is the one on your machine at two in the morning, failing on the second turn of a conversation in a way the documentation says is impossible. Ryan advises on the second one. That is where your money is actually going.

He built the thing the consultants demo

There is a multi-agent system running on his desk right now. Twelve agents sharing one memory store, routing live across eight model providers, some hosted, some running on hardware he owns, switching providers mid-conversation when one of them misbehaves. At the centre sits a memory architecture he invented, now a pending UK patent, number 2604079.0.

It has outages. It has quirks. When it breaks, and it does, there is no vendor to call. He opens the logs and finds out why, because nobody else is coming to do it.

He is not a developer and he does not sell software, and that turns out to be the qualification. He is the user every AI company claims to understand, running a system more complex than the ones their sales teams demo, writing down what actually happens. What follows is that knowledge, applied to your setup.

The thing almost nobody checks

Documentation describes how a system is meant to behave. The two routinely disagree. A provider advertises an OpenAI-compatible endpoint and then fails on the second turn of a conversation. A trace that was present on turn one is missing on turn two. A key expires and every caller upstream reports it as a broken feature, so the fault looks like your software when it is not. He has hit each of those on his own system, found them, and fixed them. Measuring that gap between promise and behaviour is the job.

What he looks at

Model selection. Which model actually does your job, at the cheapest tier that does it. Most systems are running the most expensive model on tasks a fraction of the cost would have handled. The fix is often a one-line change.

Local or cloud. What genuinely runs on hardware you own, what never will, and where the crossover sits once you have counted electricity, your own time, and the licence.

Routing and resilience. What happens when a provider fails, throttles you, or quietly stops working. Most systems have no second choice, and have never tested the one they think they have.

Cost. Where the money is going and which configuration does the same work for less. Sometimes the honest answer is that you should be spending twenty pounds a month instead of twenty thousand a year.

Problems that span paradigms. The ones that do not fit inside one vendor's product. If your problem runs across a local model, a hosted model and a memory layer, no single provider will solve it for you, because it is not in their interest to.

Now the part where the headline deflates

If the right answer is that you do not need any of this, you will be told exactly that, in writing, and it will cost you a fraction of finding out the other way. He takes no commission, margin or referral fee from any provider, so there is no kickback hidden inside the recommendation. He will not write or maintain your software. He will not resell, host or take a margin on anything he recommends. And he will not sell you a document that says adopt AI.

How it works

One engagement, fixed scope, a fee agreed in writing before anything begins. You get a written review with recommendations you can hand to whoever builds things, and a call to go through it. If you only need one decision made, that can be a single call instead.

To start, email with what you are trying to do and what you are currently spending.

ryanmusic@mac.com ↗