We make the data models learn from
Manthano AI is a Swiss company built around one job: turning raw data into training data a model can actually be trusted to learn from.
Almost every failure we are called in to look at is a data failure wearing a model costume. The label set did not match the question, the guideline was ambiguous so two annotators answered it differently, or the dataset was collected once and never audited again. The work we do sits at that layer, on purpose, because it is the layer where the damage is cheapest to undo.
Where the name comes from
Manthanō (μανθάνω) is ancient Greek for “I learn”, and more precisely for learning by observation: coming to know a thing by watching it rather than by being told. The same root gives μάθημα, máthēma, “that which is learnt”, which is where mathematics and the German Mathematik come from, and μαθητής, mathētḗs, the one doing the learning.
It is an accurate description of supervised machine learning, which is a system observing many examples until it can generalise from them, and it is an accurate description of the job. A model learns from what it is shown. We are in the business of what it is shown.
What we do
Four services, sold separately, and most projects only ever need one.
- Data annotation and model feedback
- Human-labelled training data across text, image, video, audio and 3D, and the feedback layer for language models: preference data, written demonstrations, rubric scoring, red-teaming.
- AI application development
- Custom AI-driven applications, taken from concept through to deployment by the team that labelled the data.
- Machine-learning data preparation
- The work before training: cleaning, structuring and checking datasets so a model has something honest to learn from.
- Synthetic data generation
- Rendered training data built in 3D instead of collected and labelled: for classes that are rare, expensive to photograph, or cannot leave the building.
They sit under one roof because the seams between them are where projects usually break. A dataset handed from one vendor to another loses the guideline that explains it, and a team that never read the raw data inherits its faults without knowing they are there. The services page goes through them properly.
What we hold to
Four working rules, written down because they are the ones that cost us something.
- No number we have not measured
- You will not find an accuracy percentage, a turnaround time, a project count or a client logo anywhere on this site, because we have not measured them for your project and a figure from someone else’s is decoration. Where an answer genuinely depends on the work, we say that it depends on the work. It reads weaker than the alternative and it is the only version we can stand behind in a meeting six months later.
- The guideline is the deliverable
- Labels are the output; the written definition of what each label means is the thing that makes them reproducible. We would rather spend the first day arguing about an edge case than discover on the last day that two annotators had been answering different questions. Whatever we agree gets written down and travels with the data.
- People do the judging, and we say who
- Annotation is human judgement at scale, so who does it is not an implementation detail. Part of our work is deliberately routed to annotators in places where this kind of income is worth disproportionately more, which is a position we hold rather than a cost trick, and it only works if the training and the guidelines are as good as the ones we would want for ourselves. Anyone can register; work is matched per project, in particular languages and domains, and we do not promise it in advance.
- Swiss, and specific about it
- We are based in Switzerland and Swiss processing is available on request rather than by default, because it depends on the tools a project needs and where the people doing the work are. Ask while you are scoping and we will confirm in writing what is possible for your data before any of it moves. “Swiss” on a marketing page is worth nothing without that sentence attached.
Where we are trying to get to
The company was founded on the belief that progress in artificial intelligence should come from many people contributing rather than from a few institutions with the largest private datasets, and that the data layer is where that is decided. Models get published. The data they were trained on usually does not.
So the ambition is unglamorous and long: be good enough at the teaching layer that high quality data is something an ordinary team can get hold of, and that the people producing it are paid properly for judgement that no amount of compute replaces.
Programmes and partners
None of these is a customer reference and we are not presenting them as one.
The Swiss federal innovation agency, which funds applied research run jointly by companies and Swiss universities.
NVIDIA's programme for AI companies, which provides technical resources and hardware support.
Also Google Cloud, Humans in the Loop, Lightly and Connect AI.
Where to find us
Rebmattli 52c, CH-8832 Wilen bei Wollerau, Switzerland. The full legal details, including the company register entry and the VAT number, are on the imprint, and what we do with the data you send us is set out in the privacy statement. For anything else, info@manthano.ai reaches a person.