Built to order

Custom AI Models

A model that knows your material, answers in your voice, and runs on hardware you own.

Yours to keepLocal or hostedNo lock in
Every model you can download is a general one. It knows a great deal about the world and nothing whatsoever about you.

That gap is the whole job. A general model can write you a polite email about a refund policy. It cannot tell a customer what your refund policy is, because it has never seen it.

We close that gap. Depending on what you need, that means wiring a model into your documents so it answers from them, giving it a role and a personality it holds onto, training it on your material, building the interface around it, or all four.

What comes out the other end is yours. It runs on your machine or your server, the material it learned from never comes to us, and there is no monthly fee to a company that could change its mind about you.

What we build

Two very different sorts of request, and we take both seriously.

Most of this work splits into things a company needs and things a person wants. The engineering underneath is largely the same. The conversation around it is not, so we have written them out separately.

For a business

  • A model that answers your website visitors properly. Wired into your products, your documentation, your pricing and your policies, so it answers from what is actually true today rather than what it half remembers.
  • An internal assistant with a job title. Trained on your company material and given a defined role, so new staff can ask it the questions they would otherwise queue up to ask a colleague.
  • Document and knowledge search that understands meaning. Ask a question in plain language, get the answer and the page it came from, across everything you hold.
  • Support that knows the difference between a real answer and a guess. Built to say it does not know and hand over, rather than invent something confident and wrong.

For a person

  • An assistant on your own PC. Trained on your files, your notes and the things you keep referring back to, with an interface built around how you actually work.
  • A specialist for what you are into. A coding partner that knows the toolchain you use, including the awkward ones such as building mods for a particular game, where the general models are vague and the documentation is scattered.
  • Your own retrieval system. Years of documents, research or writing, searchable by meaning, sitting entirely on your machine.
  • A personal AI buddy. A character with a personality, a memory of your conversations and a manner you choose. Running locally, which is the only arrangement under which something like that is genuinely private.
The part that matters

We ask what it needs to do.
Not why you want it.

We will ask plenty of questions, because we cannot build the right thing without knowing exactly what it is for. What we will not do is have an opinion about it. Tell us what it needs to do and we will tell you how to do it properly, and once it is handed over there is nobody at this end watching how you use it. A private assistant that reports back to its maker is not private, whatever it says on the box.

Three ways to build one

Most people ask for training. Most people need retrieval.

There are three levers, they cost wildly different amounts, and picking the wrong one is the most common way this sort of project goes wrong. We will tell you which you need, including when the answer is the cheap one.

Usually this

Retrieval

The model is given a searchable copy of your material and looks things up before it answers. Fast to build, cheap to change, and it cites where the answer came from. Update a document and the model is current the same afternoon, with no retraining at all.

Sometimes this

Role and behaviour

Shaping how the model conducts itself: what it is, how it speaks, what it refuses, what it does when it is unsure. Cheap, quick, and responsible for far more of the finished feel than people expect.

Occasionally this

Training

Actually adjusting the model's weights on your material. The right answer when you need a genuinely different style or a specialism no amount of looking things up will give you. Slower, costlier, and worth it when it is worth it.

Retrieval beats training for facts

If the problem is that the model does not know something, teaching it is the expensive way to fix a lookup. Training changes how a model behaves. Retrieval changes what it knows. Confusing the two burns budgets.

The interface is half the product

A brilliant model behind a bad window gets abandoned in a fortnight. We build the thing you actually click on as well, whether that is a panel on your website or a proper application on your desktop.

The base model matters less than you think

Once it is wired into your material and given a role, the difference between two decent open models is usually smaller than the difference between two ways of preparing your documents. More on choosing one.

Where it runs

On your hardware, or on infrastructure we manage for you.

This is a sovereignty company, so local is never the awkward option. For anything personal it is the obvious answer. For a system serving customers around the clock, managed hosting usually is, and we will say which is which rather than selling you the one that sounds better.

Your machine

  • The whole thing sits on the desk. Model, documents, conversations, all of it, with no connection required to use it.
  • Nothing you feed it is transmitted anywhere, because there is nowhere for it to go.
  • It keeps working regardless of us. No licence check, no server to call, no deprecation notice.

Your server, or hosted

  • On your own hardware when the whole team needs to reach the same assistant from inside your own network.
  • Or hosted and managed, which is how most website facing systems run, because something answering your customers at three in the morning needs to be up when the office is not.
  • We will tell you which of the two your situation actually calls for. More for businesses.

What we keep

  • Your material is working stock, not ours. We use it to build the thing and we do not train anything else on it.
  • Nothing feeds back to us in use. There is no telemetry channel from a finished build to this office.
  • You own what we hand over. The model, the setup and the interface are yours to keep, move, or take somewhere else entirely.
Before you ask

What a custom model will and will not do for you.

We would rather talk you out of the wrong build than take the money and hand you a disappointment.

Some of these come up in almost every first conversation, so here they are in advance.

Realistic
  • Answering accurately from your material. Every answer carries a reference to where it came from.
  • Holding a role and a manner. Consistently, across long conversations and between separate sessions.
  • Deep familiarity with a narrow subject. Including the ones general models handle badly.
  • Running properly on ordinary hardware. Sized to the machine you actually have.
Not realistic
  • A model that never gets anything wrong. They are not built that way. What we can do is make it show its source and admit when it does not know, so you can tell.
  • Frontier ability on a laptop. A local model is bounded by the hardware under it, and no amount of tuning gets round that.
  • Teaching it facts by training alone. That is what retrieval is for, and we will say so.
How it goes
  • A conversation first. What it is for, who uses it, what hardware exists, what "working" looks like to you.
  • Your material, in whatever state it is in. Scattered, inconsistent and half in PDFs is normal, and tidying it is part of the job.
  • A build you try, then adjust. Nobody describes what they want perfectly first time, and we do not expect you to.
  • Handed over properly. Set up on your hardware, with instructions written for a person rather than an engineer.
Start a conversation

Tell us what it needs to do.

No form, no discovery call, no funnel. Describe the thing you want in whatever words you have for it, tell us what machine it has to run on, and we will tell you honestly whether it is a weekend of work or a serious project.

Enquiring for a business? Our Enterprise page covers deployment, hosting and managed systems at scale.