Custom AI development and model integration

When you need something built rather than configured

Most AI problems are solved by connecting a good existing model to your business properly, and that's what our AI assistants work covers. This page is for the cases where that isn't enough: the AI has to live inside your product, reason over data nobody else has, or make a judgement specific to your industry that no off-the-shelf tool understands.

That's a different kind of project, with a different risk profile. It can be genuinely transformative, and it can also burn twenty thousand pounds producing a demo that never survives contact with real data. The way you avoid the second outcome is to test the idea cheaply, on your actual data, before committing to the build — which is why every engagement here starts with a paid discovery rather than a proposal.

What we build

• Document and paperwork processing. Invoices, purchase orders, certificates, survey reports, timesheets — read, checked and turned into structured data in your systems instead of typed in by a person.
• Search and answers over your own material. Years of manuals, job records, contracts and specifications made answerable in plain English, with citations back to the source document so answers can be checked.
• Classification and routing at volume. Thousands of items a day sorted, scored, prioritised or flagged for review by rules that would take a person a week.
• Forecasting and prediction. Demand, staffing, stock, seasonal patterns, churn risk — trained on your history rather than a generic industry assumption.
• Image and photo analysis. Condition assessment, damage triage, automatic measurement and quality checks from photos your staff or customers already take.
• AI features inside your own product. If you sell software, this is the AI capability your customers are asking for — built, evaluated and shipped as part of the product rather than bolted on.

How we work

1. Free framing call. What decision or task are you trying to improve, and what data exists about it? A surprising number of ideas die well here, at no cost, because the honest answer is a rule or a spreadsheet.
2. Paid discovery — £950, one to two weeks. We take a sample of your real data, build a rough version, and measure it. You get a written verdict: what accuracy is achievable, what it would cost to build properly, and whether it's worth doing. If the answer is no, that's a successful discovery.
3. Evaluation first. Before the real build we assemble a scored test set from your data. Every change gets measured against it, so "better" is a number rather than an opinion.
4. Build and integrate. The model is the small part. Getting data in and results back out where your team actually works is the integration work, and it's typically most of the project.
5. Human review by default. AI output that affects money, safety or commitments goes past a person, with the confidence score visible so obvious cases move fast and marginal ones get looked at.
6. Monitor and maintain. Accuracy drifts as your data and the underlying models change. We watch the live numbers and tell you when something has slipped, which is not optional for anything you depend on.

What it costs

Discovery is £950, fixed. Builds start at £4,500 for a single well-scoped capability and run to £25,000 and above for AI embedded in a product you sell. Ongoing model usage, hosting and monitoring generally land between £150 and £600 a month depending on volume, and we estimate it properly at proposal stage rather than leaving you to find out.

You own everything: code, pipelines, prompts, evaluation sets, any fine-tuned weights and all of the data. As with the rest of what we do, there's no lock-in and no arrangement where walking away means starting over.