AI data and models
Models trained on
your ground truth.
A model that has never seen your environment will keep telling you about somebody else's. We build, train and evaluate models on the data you already hold, inside the boundary you already trust.
What we build
Data foundation
Labelling, ground truth, and the unglamorous work of knowing what your data actually says before any model is allowed to learn from it.
Training in place
Training runs where your data lives: your region, your tenancy, or entirely inside your perimeter. The corpus does not travel.
Evaluation against your adversaries
Measured on the threats you face, with false-positive rate reported as a number on a named corpus rather than as a claim.
Governance
Versioned models, recorded thresholds, and decisions that can be replayed months later against the weights that produced them.
The standard
An uncalibrated score is never allowed to act alone.
A model may recommend. Whether it is permitted to act, and how far, is a policy decision with a written boundary, and that boundary is enforced in the product rather than described in a slide.
The corpora
The measurement is the asset.
Corxor maintains standing labelled corpora — adversarial and benign, in several languages — built and curated over years rather than scraped. They are versioned, disputed internally, and grown every time the field produces something the existing set does not cover.
No release ships until it has been re-measured against them. A build whose numbers move in the wrong direction does not go out, regardless of what else is in it.
The model family that sits on top of them — PromptIntent among them — is asked for on its own, independently of any endpoint deployment.
Base-tier access to that family can be extended to institutions we have accepted, under contract, and kept entirely separate from the models trained on their own data.