AI
AI services, training data and evaluation in Portuguese
Training data, annotation, evaluation and safety, agents and document intelligence, with specific attention to European Portuguese and the African variants almost nobody covers well.
Why is Portuguese a specific problem for AI?
Because most available Portuguese data is Brazilian, and a model trained mainly on PT-BR produces answers that sound wrong to a Portuguese user. The problem is not comprehension, it is register: the system replies with a familiarity the Portuguese customer did not ask for and reads as inappropriate.
Add that the Angolan, Mozambican and Cape Verdean variants are barely represented in public datasets, and that evaluating output in those variants requires evaluators from those variants rather than merely Portuguese speakers. It is work few suppliers do, and the Lusophone corridor is what lets us do it.
How is annotation quality assured?
With guidelines calibrated before volume rather than during it. A small initial batch is annotated by several annotators in parallel, disagreements are discussed, and the guidelines are rewritten until agreement stabilises. Only then does volume begin.
Inter-annotator agreement is measured and reported throughout the project, not just at the start. When it falls, that usually means the data has changed in character and the guidelines need revision, and it is better to find that out from the metric than from model performance months later.
What does human in the loop mean in practice?
It means defined review points, with written criteria on when the system may decide alone and when it may not. In document intelligence, for instance, extraction runs automatically above a confidence threshold and goes to human review below it.
The common failure is treating human review as a generic safety net and ending up with people approving everything without looking. A well designed loop sends for review what is worth reviewing, and measures whether the review is changing anything.
The AI services we run
Each service has its own page, with concrete deliverables and an indication of where the variant decision applies.
Service index
- AI managed servicesOngoing operation of AI-assisted workflows, with people in the loop where the risk requires it.
- Training dataCollection, annotation and review of model training data, including European Portuguese and the African variants.Variant
- Agents and automationAutomation of repetitive processes and assisted agents, with supervision and defined limits.
- Evaluation and safetyModel output evaluation, adversarial testing and safety review, by native speakers.Variant
- Document intelligenceExtraction and classification of information from documents, with human verification where it matters.
- AI in PortugueseAI work that treats Portuguese as two or more operations, because that is what it is.Variant
Frequently asked questions
Do you work with confidential data?
Yes, with a data processing agreement, named and logged access, and an EEA-only delivery option where the data category requires it.
Do you use our data to train other models?
No. A client's data is used only for that client's work, and that is in the contract rather than in a statement of intent.
Which languages do you cover for training data?
Portuguese in the European, Brazilian, Angolan, Mozambican and Cape Verdean variants, English, Spanish, French and more than thirty others across the group network.
Can you evaluate a model we did not train with you?
Yes, and it is a frequent request. Independent evaluation is most useful precisely when the evaluator did not build the thing being evaluated.
How do you handle distressing content in moderation and evaluation?
With exposure limits, task rotation and wellbeing support planned from the outset. It is a cost of the service and is included in the price rather than stripped out to lower the bid.
What is the minimum volume for an annotation project?
There is no hard minimum, but below a few weeks of work the initial calibration weighs too heavily on total cost, and it is worth saying so in advance.
Do you do European Portuguese work specifically?
Yes, and it is one of the reasons this operation exists. Datasets separated by variant, with provenance metadata, and evaluators native to the variant assessed.
How does an AI project with you start?
With a small paid pilot batch, which produces calibrated guidelines and an agreement measure before any volume commitment.
Let us look at the numbers for your case
Tell us which processes you want to outsource, in which languages and at what volume. We come back with a euro estimate and an operating design, with no commitment.
We reply within 6 hours on working days. If you would rather write: info@corpshore.solutions