Try the demo · Download the Studio · Home
What is MatrixAI Studio?
MatrixAI Studio is a declarative, auditable machine learning framework launched in June 2026. You give it a table of data and say what you want to predict; it trains competing models, measures them honestly on data they never saw, and tells you which one to use and how far to trust it.
What sets it apart from tools like PyTorch or TensorFlow is that it does not stop at a number. Every result comes with the evidence behind it: the data used, how the models were compared, the risks it found, and a verifiable record you can show an auditor.
It is built for IT teams, AI project leads and developers who need reliable decision systems in regulated or sensitive settings — where it is not enough for a model to work: you need to prove why it works and when it might fail.
What it does today
Models that compete
Upload a CSV, pick the column to predict and start a study. Gradient-boosted trees, neural networks (including TabM, a state-of-the-art tabular network) and a plain reference baseline all compete on the same data, under the same rules. The winner is chosen by cross-validation and then confirmed on a held-out test set that was never used to choose. Classification (two or more classes) and regression are supported.
Honest results
You see the comparison, not just the winner: how much better it is than doing nothing, how sure we are of that, the risks found in your data, and the cost of getting it wrong if you tell us what a false alarm and a miss cost you. Where the data does not support a conclusion, the Studio says so instead of drawing a pretty chart.
A laboratory to see a network learn
For people who want to understand, not just use: the laboratory trains a network on your table and lets you look inside — the weights of each layer, the path one row takes through the network, how each layer learns. Change one setting at a time with a fixed seed, run it again and compare the runs side by side.
Evidence you can verify
Every study produces a record of its configuration, data and results, signed when a signing key is configured. The saved package can be re-checked later: integrity, recomputed metrics, repeated inference. Nothing happens without leaving a trace.
Declarative language (.mxai)
Under the hood, models are described in .mxai and .mxtrain files, not arbitrary Python. Shapes, types and ranges are verified before any training starts, so incompatibilities surface at compile time, not in production. Every model can be inspected, versioned and reproduced.
Yours to keep and to serve
The Studio you download runs on your own machine: your data does not leave it. Take the trained model as a package or a Docker service; the expert mode also exports ONNX/WASM and edge bundles.
Frequently asked questions
Who is it designed for?
For development teams, AI project leads and organisations that need explainable decision systems in healthcare, finance, logistics or any setting where auditability is a requirement.
Is MatrixAI Studio open source?
The core is AGPL v3, published on GitHub and PyPI as matrixai-core. The Studio is a commercial package you download and run on your own machine.
How is it different from other AutoML tools?
It reports how much to trust the result, not only the best score: the held-out test, the risks in the data, the cost of errors and a verifiable record of how the choice was made.
What audit guarantees does it offer?
Every study and every training run produces a trace (signed when a signing key is configured) with the configuration, the data used, the metrics and the fingerprint of the resulting model. Production decisions can be recorded too.
Can I try it without installing anything?
Yes. The online demo runs the full cycle in your browser with guided examples, from the data to a trained, compared model — nothing to install. Uploading your own CSV is available in the downloadable Studio. Back to the home page.
Technology stack
- Frontend: React, TypeScript, Vite
- Backend and core: Python, own type system, declarative compiler
- Models: LightGBM, scikit-learn, PyTorch (TabM and dense networks)
- Infrastructure: Docker, Nginx, SSL/TLS
- Export: ONNX, WASM, HTTP/REST with Bearer authentication
Contact
For collaborations, technical questions or commercial enquiries, visit our contact page.