Skip to content
BlogPublished 17 November 20265 min read

Software Development United Kingdom Without the Sales Pitch

AI productssoftware developmentremote engagementMVP sprintverification

When a founder starts comparing partners for software development United Kingdom, they usually get a slide deck and a promise. I run a studio in Douala and I have shipped AI products for clients in London, Berlin, and Zürich. This note is the opposite of that slide deck. It is what actually happens when you hire me to build an AI product, from the first call to the day the model goes live.

The First Call Is About the Workflow, Not the Model

Most founders come to me with a model in mind. They have read about large language models or vision APIs and they want to put one in their product. I ask them to slow down. The first conversation is not about which model to use. It is about the workflow the model will sit inside.

For Fursa, the visa route eligibility tool, the client came with a clear problem: people waste hours reading immigration rules and still get the answer wrong. The model was a means to an end. We spent the first call mapping the exact steps a user takes, from entering their nationality to seeing a list of possible visa routes. Only after that did we discuss how a model could compress those steps. If you cannot describe the workflow on a whiteboard, the model will not save you.

Software Development United Kingdom Starts With a Constraint

Every AI product I have built shares one constraint: the model output has to be right, and right is defined by something outside the model. For Fursa, that something was the official immigration guidance from 170 destination countries. The model could not hallucinate a visa route. It had to cite a source for every claim. For Job Hunter, the constraint was freshness. A job board that shows yesterday's listings is useless, so the model's crawl and match step had to be verified against the career pages themselves.

This constraint changes how I run the engagement. I do not start with a prompt and a playground. I start with a verification design. Who checks the model's answer? What source does the checker compare against? What happens when the checker disagrees? Those questions get answered before a single line of model integration code is written. When a UK client hires me for software development United Kingdom, they are not buying a chatbot. They are buying a system where the model is one component and the human gate is another.

The First Sprint Exposes the Real Problem

I run most AI product builds as an MVP Sprint. Eight to twelve weeks, fixed scope, one thin slice of the workflow. For ProPost AI, the LinkedIn presence engine, the first sprint was not about generating 365 drafts per user. It was about generating one draft that matched a user's voice and letting that user approve or reject it before anything else happened.

The sprint exposed a problem the client had not anticipated. The model could produce fluent text, but it could not tell the difference between a professional update and a humble brag. We spent three days building a scoring layer that checked tone against a set of examples the client provided. That layer was not in the original scope. It became the core of the product. If we had skipped the sprint and gone straight to a full build, we would have shipped a generator that nobody trusted.

Verification Is a Feature, Not a Phase

In every AI product I have built, verification is not a step at the end. It is a feature that runs in the product. The user sees the model's output and the evidence behind it. The user can approve, reject, or edit. The system logs every decision.

For GritGateway, the African talent intelligence platform, this meant a candidate profile was never shown to an employer without a human reviewer confirming the skills matched the source data. The model extracted skills from CVs and public profiles. A reviewer saw the extraction side by side with the original. One click approved it. One click sent it back. That loop ran hundreds of times a day and it kept the false positive rate low enough for recruiters to trust the platform.

Here is the verification loop I use on every AI product:

  • The model produces a candidate answer with a confidence score.
  • A source document or external API is fetched and compared.
  • A human reviewer sees both and approves or rejects.
  • The decision is logged and fed back into the model's evaluation set.

This is not a research project. It is a production system. The human approval step is not a temporary patch. It is the reason the product can be sold to a regulated client or a public sector body.

Shipping Means Owning the Outcome

When the sprint ends, the product does not vanish. I stay on for the launch and the first few weeks of live traffic. For a UK client, that means overlapping time zones. My day in Douala starts early, so I have four to five hours of overlap with London. I use that window for pairing, for reviewing logs, and for walking the client through the verification dashboard.

Shipping an AI product is not the same as shipping a CRUD app. The model will behave differently in production because real users ask things you did not anticipate. In the first week after launch, I watch the rejection rate. If it spikes, we adjust the prompt, the source set, or the approval threshold. The client sees the change and the effect in the same day. That is the difference between a partner and a vendor. A vendor hands over code. A partner owns the outcome until the rejection rate settles.

The Conversation That Should Come Next

If you are comparing partners for an AI product, do not ask for a portfolio. Ask for a constraint. Ask how the last project verified a model's output before a user saw it. Ask what happened when the verification failed. I can answer those questions with the cases I have linked here, and I can answer more specific ones in a call.

Start that conversation through the contact form or ask Kadon AI a question from my record. The worst outcome is that you learn something about your own workflow before you spend a pound on a model.

Want to talk about something here?

Let’s talk about it.

Start a conversation