Our model
Match recognises the same problem, whatever the words.
No two customers describe a problem the same way. Match is the model we trained for a single question: are these two pieces of feedback about the same problem? It is why a signal gathers everything that belongs to it, and nothing that doesn't.
Match 1.0 · in service since 2 October 2026
- 99% of decisions right when it is sure
- FR · EN the same problem recognised across languages
- 0 pieces of customer data used to train it
When it's sure, it decides. When it isn't, it gets a second opinion.
Match never guesses. It knows the cases it is certain about from the ones worth checking.
It compares
Every new piece of feedback is set against the closest signals in your workspace. For each pair, Match estimates whether it is the same problem.
Sure, it decides
The feedback joins its signal, or stays apart. In those cases Match is right 99 times out of 100 on our test bench.
Unsure, a second opinion
The pair goes to a second, slower AI model that examines it in detail. No feedback is ever filed at random.
A model of our own, so your feedback stays yours.
Trained without a single piece of customer data
Match learned from tens of thousands of examples we created ourselves, in French and English. Your feedback never trains it.
Hosted in France
It runs on a server dedicated to Cleansig, at OVHcloud. The sentences it compares never leave the European Union, and it keeps none of them.
Measured before every version
Every version passes our test bench before it goes into service. In production, a check confirms it makes the same decisions it made on the bench.
Every version, dated and measured.
A new version of Match only replaces the previous one once it does better on our test bench. The old one stays ready: we can roll back in an instant.
Match 1.0 · 2 October 2026
- First version of Match, trained on tens of thousands of pairs of product feedback, in French and English.
- Decides alone when it is sure, and hands the rest to a second AI model.
- 99% of decisions right when it is sure, on our test bench.
Precision measured in October 2026 on our internal test bench: product feedback in French and English, written to trap the model with closely related problems.
What people ask us about Match.
What is Match?
It is the model Cleansig trained for a single question: do two pieces of feedback describe the same problem? It recognises the problem even when your customers put it in other words, or in another language. It is why a signal gathers everything that belongs to it, and nothing that doesn't.
What does “right 99 times out of 100 when it's sure” mean?
For every comparison, Match knows whether it is sure. When it is, it decides alone, and on our test bench it is then right 99 times out of 100. When it isn't, it doesn't guess: the comparison goes to a second AI model that examines it in detail.
Who gives the second opinion?
A second, slower and more thorough AI model, run by Mistral AI within the European Union. Like all our providers, Mistral AI is contractually committed not to train its models on your content.
What does Match see of my data, and what does it keep?
Only the signal statements it has to compare, never your whole meetings. It runs on a server dedicated to Cleansig, at OVHcloud, in France: it compares the sentences and keeps none of them. And it was trained without a single piece of customer data.
Does Match understand several languages?
It was trained in French and English, and recognises the same problem across the two: feedback from an English-speaking customer joins the signal a French-speaking customer opened.
What if Match is unavailable?
Cleansig then hands every comparison to the second model, as it did before Match. No feedback is lost or filed at random: it is just a little slower.