Skip to content
Intura
enid
Enterprise Small AI Models: Custom Training and Integration / Custom models
Develop

A private model, trained on the context only your company has.

Hand us the decision. We train a model on your labels, compress it until it runs on a CPU, deploy it where your data already lives and build the pipeline that keeps it current. Then we hand it over as something your team owns and can run without us.

The problem

The problem it solves

A catalogue can only cover what looks the same everywhere. The decisions that determine whether your system works run on categories your operations team invented, a policy only your risk function wrote, a language mix no public corpus contains, and years of your own case history. A general model prompted with your policy gets close and stays close; a model trained on that context learns it. And a model that is never retrained decays against the people it is meant to stop, who rewrite their approach every week.

Who it is for

Who it is for

  • Risk, fraud and trust-and-safety teams with their own taxonomy and case history

  • Operations teams classifying, routing or extracting from documents and messages in volume

  • Companies serving Indonesian, English or a mix of both, where general models handle the language poorly

  • Teams whose data cannot leave their perimeter, for training or for serving

What we do

What we do

  1. 1

    Define the decision

    Your categories, your policy wording and your escalation rules, written down as the labels the model will learn, with your team agreeing where each boundary sits.

  2. 2

    Build the dataset

    Ingestion, cleaning and labelling on your own cases, versioned, with checks on class balance and quality. Where your policy requires it, this happens inside your environment.

  3. 3

    Choose the model

    Classical machine learning, a small language model or a fine-tuned LLM, chosen for the decision rather than for the brochure. The smallest one that does the job is often the right answer.

  4. 4

    Train and compress

    Trained on your labels, then compressed: vocabulary trimmed to the languages you serve, INT8 quantisation, and the model re-checked against your evaluation set after each step so that small never quietly means worse.

  5. 5

    Deploy and keep current

    Deployed in your VPC, on-premise, air-gapped or on the device, with drift monitoring and a retraining path that ships a new version as a file swap rather than a rebuild.

What you get

What you take delivery of

  • The model, owned by you, with the evaluation set it was checked against
  • A small artifact, quantised and benchmarked, that runs where you need it
  • The dataset, the labelling guide and the training and evaluation pipeline
  • Deployment in the environment you chose, with a runbook
  • A retraining path your team can run without us

How it runs

How an engagement runs

A custom model usually takes longer than an integration, and most of the time goes to the data rather than to training: agreeing the labels and collecting enough well-labelled cases. The proposal after the first call states the timeline for your case and what we need from your team to keep to it.

  1. 01

    First call

    Tell us what the model has to decide. You get an engineer, not a qualification script.

  2. 02

    Scoping and data review

    We look at a sample of your data and your policy, and write down the labels, the baseline to beat and the environment.

  3. 03

    Dataset, training and compression

    Built in iterations, with a review of the model's mistakes on your own cases at each one.

  4. 04

    Deployment and handover

    Installed where it will run, benchmarked on that hardware, handed over with the pipeline and the runbook.

FAQ

Common questions

Do you need a GPU to train a custom model?

Not one you own. Training is our side of the engagement; the model you run afterwards is CPU-only, like everything else on this site. If your policy requires training to happen inside your own environment, say so early. It is a normal requirement and we scope for it.

Can you build a model small enough to run on the edge?

That is the specialisation. Every model on this site was made small on purpose: the vocabulary trimmed to the languages served, layers cut, the graph quantised to INT8, and the model re-checked after each step rather than assumed to be fine. We do the same for a custom model, and tell you what each compression step cost.

Who owns the model and the data?

You do. The model, the dataset, the labels and the pipeline are delivered to you, and none of it is a licence you keep paying for.

Indonesian, English, or both?

Either or both. As in the catalogue, each language is usually its own build with the vocabulary trimmed to it, so a bilingual service gets two small models and a step that sends each message to the right one.

See what a model page publishes

Every catalogue model shows its base model, what it was trained on and real example results. A custom model comes with the same documentation.

Browse the models

The other Enterprise services

Enquire

Tell us what you need

Describe the problem, the team and where it has to run. An engineer reads every enquiry and replies by email.

You will get an engineer on the first call, not a qualification script.