Service · AI and ML engineering

Custom model building

A model trained on your own data, for one decision your business makes every day.

Relevant to Manufacturing, Service businesses, Education, Growing businesses

The problem

Off-the-shelf tools do not know your product codes, your seasonality or your failure modes. A general model gives general answers.

You have a decision that is made many times a day, by a few experienced people, from the same inputs: which enquiry to call, which batch to hold, which file to look at twice. That is what a model is for.

What it costs today

Typical: a decision made hundreds of times a day whose quality depends on who is on shift, with no record of how often it was right.

This is for you if
  • One decision, made repeatedly, with a recorded outcome: the order came or did not, the part failed or did not, the student stayed or left.
  • A few thousand past examples, or a few hundred with a strong signal.
  • Someone on your side who makes that decision today and will tell us when the model is wrong.

Probably not for you if

  • A decision made twenty times a year. A checklist is cheaper and better.
  • A problem where nobody can say what a right answer looks like. We help you define it first, in the assessment, before anyone trains anything.
What we build
  • A model for one named decision: a forecast, a risk score, a classification, a ranking, or a reading from an image or a document.
  • A measured accuracy on data the model never saw, compared against how the decision is made today.
  • A plain explanation for each output: which inputs pushed it, in words a supervisor accepts.
  • Deployment as a small service inside your systems, with a retraining schedule.
How we build itIn this order
  1. 01

    Define the decision and the baseline

    We write the decision as one sentence and measure how often today's method gets it right. That baseline is the number the model has to beat.

  2. 02

    Build the training set from your records

    Linking inputs to outcomes across systems is most of the work. We hold out a later period for testing, so the accuracy we report is honest.

  3. 03

    Start simple

    A gradient-boosted model on tabular data wins most business problems. Deep learning is used for images, text and signals, and only after the simple model has been beaten.

  4. 04

    Explain, then deploy

    Every prediction carries its reasons. The model runs as an API or a batch job in your environment, and shadows the current method for a few weeks before anyone acts on it.

  5. 05

    Monitor and retrain

    Accuracy is published every cycle. Retraining runs on a schedule, and the model is retired if the business changes under it.

Data you already haveThis is usually enough to start
  • Records of past decisions and what happened afterwards: orders, defects, defaults, dropouts, returns
  • The inputs available at decision time, not after it
  • Images, documents or sensor signals, where the decision is read from them
  • Who made each decision, if you want to compare people to the model
In practice

Illustrative example

Quoting job work in an hour instead of two days, with a cycle-time model

Every quote depended on one senior estimator reading the drawing and guessing the cycle time. Quotes took one to two days, and the factory lost orders to faster competitors. When the guess was low, the job ran at a loss; when it was high, the order went elsewhere.

Read the case study
Typical engagement
Duration2-week assessment, then 8 to 12 weeks to a shadow deployment
TeamML engineer, data engineer, your decision owner
You receiveDeployed model, measured accuracy against the baseline, model card, retraining job
PriceFixed price, quoted after the readiness assessment
What you own at the end
  • Source code, training pipeline, model artefacts and the held-out test set, in your repository.
  • A model card: what it predicts, on what data, with what accuracy, and where it is weak.
  • The retraining job and a monitoring page your IT team can read.

Everything runs in your environment and is handed over with documentation. Nothing stops working if we part ways.

Questions we are usually asked

Do you use the latest large language models?

Where the input is text or documents, often yes, as one component. For most tabular business predictions a smaller model trained on your data is more accurate, cheaper and easier to explain. We choose after the assessment, and we show you the comparison.

How accurate will it be?

We do not know until we have your data, and we distrust anyone who says otherwise. The assessment gives a first estimate from a sample, and the pilot measures it on data the model never saw.

Who owns the model?

You do. Code, weights and training data stay in your environment, and the handover includes the retraining job.

What if our data is too small?

We say so in the assessment and propose a rule-based system that collects the right data from day one. A model follows when the data supports it.

Can it run without internet?

Yes. Models we build run on a server inside your network if that is what you need.

Often combined with

Further reading

Not sure if your data is ready?

Take the free 20-minute readiness check. You get a short scorecard and a plain answer about what, if anything, is worth doing next.

Start the readiness checkTalk to us