Machine learning

What is it?

Custom machine learning means training a predictive or classification model specifically on your historical data. Unlike generic LLMs, an ML model trained on your data is optimised for your exact use cases and runs without any external API dependency.

How it works

1. 1

   Data analysis and preparation

   Exploring available data, assessing quality, cleaning and building relevant features for the model.

2. 2

   Model selection

   Comparing approaches (regression, classification, clustering, NLP) and choosing the model based on constraints: performance, interpretability, inference cost.

3. 3

   Training and evaluation

   Training with cross-validation, hyperparameter optimisation and evaluation on a held-out test set. Metrics provided according to the use case.

4. 4

   Deployment and integration

   Model packaging (REST API or batch), integration into your existing systems and production performance monitoring.

5. 5

   Maintenance

   Model onboarding support in production, metric tracking and adjustments if data distributions drift.

What it covers

• Model trained specifically on your data to outperform generic solutions
• Data preparation, training, evaluation and deployment handled end-to-end
• Use cases: classification, extraction, enrichment, semantic search
• Ideal for companies that already have historical data to leverage

Related projects

• Automatic business classification - Médée

Frequently asked questions

How much data do I need to train a model?

▾

It depends strongly on the problem. For simple binary classification, a few hundred labelled examples are often enough. For NLP or complex cases, a few thousand are preferable. The audit phase assesses this before any commitment.

What is the difference between machine learning and an LLM?

▾

An LLM is a general-purpose language model pre-trained on billions of texts. A custom ML model is trained only on your data for a specific objective (predicting churn, classifying a document). Custom ML is often faster, cheaper at inference and more accurate on specific business tasks.

Can the model be updated as my data evolves?

▾

Yes. Delivery includes documented training and validation scripts so your team can retrain the model on new data. One-off retraining engagements are also available.

Is the ML model GDPR-compliant?

▾

If the model is trained on personal data, a Data Protection Impact Assessment (DPIA) may be required under Article 35 of the GDPR. On-premise or EU cloud deployment ensures that inference data does not leave the EU. These points are evaluated during the initial audit.