How our Artificial Intelligence projects work

Discovery call

We get on a video call (or meet in person if you are in south Wales) and you walk us through the problem. What decision are you trying to improve? What data do you have? What does success look like in numbers? We take notes, ask questions, and tell you honestly whether AI is likely to help or whether a simpler solution would serve you better.

There is no charge for this call. If we think the project is not a good fit, we will say so and, where possible, point you toward someone who can help.

Data review and scoping

You share a sample of your data (we can sign an NDA first). We profile it: row counts, missing values, class distributions, date ranges, join keys. From this we write a one-page scope document that covers the proposed approach, the metric we will optimise for, the deliverables, the timeline, and the fixed price.

If your data turns out to be too sparse or too noisy for the question you want to answer, we flag that here rather than letting you pay for a model that would not perform. We have declined about one in five enquiries at this stage because the data was not ready.

Data preparation

This is usually the longest phase. We clean the dataset, handle missing values, encode categorical features, and engineer new features where they add predictive power. For NLP projects, this includes tokenisation, stop-word removal, and building a domain-specific vocabulary.

You receive a data-quality report at the end of this phase. It documents every transformation we applied and why, so your internal team can reproduce the pipeline later without our help.

Model training and evaluation

We train multiple candidate models and compare them on a held-out test set using the metric agreed in the scope document (for example, F1 score for a classifier or mean absolute error for a regression). We also run fairness checks to make sure the model does not systematically disadvantage any subgroup in your customer base.

At the end of this phase we present the results in a short slide deck: what each model scored, which one we recommend, and what its known limitations are. If none of the models beat the baseline by a meaningful margin, we tell you that too.

Deployment

We package the chosen model and deploy it into your environment. That might mean a REST API endpoint on AWS, a scheduled batch job that writes predictions into your database every morning, or an Excel add-in if your team lives in spreadsheets. We test the integration end-to-end with real data before handing over.

Handover and support

You get a documentation pack that includes a plain-English summary of what the model does, a technical reference for your developers, and a recorded video walkthrough. For the next 60 days, you can email us with questions or bug reports and we will respond within one working day.

If the model drifts (accuracy drops because the underlying data patterns change), we retrain it once during this support window at no extra cost. After 60 days, you can extend support via a monthly retainer or manage the model yourself using the documentation we provided.