Machine Learning Development: From Data to a Working Model

Machine learning development means training a model on a business’s own data to make a repeated prediction or decision: which orders will be late, which customers will leave, what a property will rent for, which transactions look wrong. GBCORP builds these inside the software the business runs, so the prediction appears where the decision is made, and treats the model as a component with an evaluation, a deployment and a maintenance plan.

Rule, commercial model, or trained model

The data a business actually has

Orders, bookings, tickets, payments, visits, with dates and outcomes. Enough of it, labelled by what happened. The scoping phase looks at the real tables, not the ideal ones, and says whether a trained model is possible yet. Often the honest advice is to start collecting the right fields now and revisit in two quarters.

Evaluation before training

The baseline is measured first: what a rule or the current process achieves. The target is written down: the error rate or lift the model must beat to be worth deploying. A held-out set is kept aside. A model that does not beat the baseline is not deployed, and the business has still learned something cheaply.

Deployment inside the software

The prediction appears in the console or the app at the moment of decision, with its confidence, and a person can override it. Overrides are recorded as new labels. The model is called behind an interface so it can be retrained or replaced without changing the software around it.

Monitoring drift

Data changes: new products, new markets, new behaviour. The evaluation runs on a schedule against recent outcomes, accuracy is reported weekly, and retraining is triggered when it slips. This is maintenance for a component that happens to learn, and GBCORP prices it as a plan.

Checklist before machine learning

• The decision, who makes it today, and how often.
• The outcome recorded in the data, and for how long.
• The baseline: what the current process achieves.
• The number the model must beat to be worth it.
• Where the prediction will be shown and who can override it.
• Who owns the data and the model: the business.

The numbers to expect

Scoping in two weeks including a data review. A first evaluated model in six to ten weeks at a fixed price, deployed behind an override. Weekly accuracy reports and scheduled retraining under the support plan. Customised or more complex functionality is quoted separately before any work starts.

A worked example

A delivery company wants to know which of tomorrow’s orders will be late, so dispatch can act before customers call. Two years of orders with outcomes exist. The baseline is the dispatcher’s instinct, measured for a fortnight. A model trained on zone, time, size, driver load and weather flags the orders most likely to be late; it beats the baseline on the held-out set by a clear margin. It is deployed in the dispatch console as a flag with a confidence, the dispatcher can override, and overrides become labels. Late deliveries fall within a quarter, and the weekly report shows the model’s accuracy holding through a seasonal change.

Where the data lives and who sees it

Training data, the model and the evaluation set live in the business’s own cloud project. Nothing is sent to a third party for training. Personal fields are removed or masked before training where the prediction does not need them. The prediction is shown to the roles that make the decision and logged with its confidence, so an audit can see what the model said and what the person did.

Retraining as maintenance

A model trained on last year’s behaviour drifts as the business changes. Under the support plan the evaluation runs on a schedule against recent outcomes, accuracy is reported weekly, and retraining is triggered when it slips below the agreed line. New fields the business starts recording are added at the next retraining. This is ordinary maintenance for a component that learns, priced as a plan rather than left to chance.

Questions to ask any machine learning company

What is the baseline and how was it measured. What number must the model beat. What is held out for evaluation. Where will the prediction be shown and who can override it. What happens when accuracy drifts. Who owns the data, the model and the evaluation set. A company that answers all six with numbers rather than slides has deployed models that businesses rely on.

Pricing that matches the risk

A data review priced on its own, so a business that does not yet have the data learns that cheaply. A first evaluated model at a fixed price, deployed only if it beats the baseline. A support plan covering scheduled evaluation, retraining and reporting. No retainer without a defined outcome. Customised or more complex functionality is quoted separately before any work starts.

Machine learning inside ordinary software

The prediction lives in the console or the app the business already uses, on GBCORP’s engine, with the same accounts, roles and audit history. A model bolted on beside the software is a report nobody opens; one inside the screen where the decision is made changes the decision.

Questions we hear first

Do we have enough data?

The data review answers that in the first fortnight; often the answer is not yet, with a list of fields to start recording.

Can we use a commercial model instead?

For language, documents and images, usually yes; for predictions from your own history, a trained model is the tool.

Who owns the model?

You do: the training data, the model and the evaluation set, in your cloud project.

What if the model is wrong?

The person overrides it, the override becomes a label, and the weekly report shows the rate.

Where to start

The GBCORP AI Development and Automation product is the priced starting point; the GBCORP AI development hub sets out the method. For models inside apps, read AI application development explained.

GBCORP operates across 17 countries on 4 continents through six pillars: AI & IT, Innovation, Sustainability, Infrastructure, Business Lounge, and EV's & Solutions.