/ platform · inferrexai
Inference you can run where your data already is.
InferrexAI is Inferrex's own inference — its own models, trained on API specifications rather than customer data, serving every deployment mode including the ones with no internet at all.
Most structural work does not need a model.
Most fields classify without AI at all. Deterministic pattern matching against The Corpus resolves them directly, and InferrexAI is reserved for the genuinely ambiguous ones.
That ordering is an architectural decision, not an optimisation. Known and deterministic paths are not forced through inference to make the platform sound more intelligent — and a system that reaches for a model first cannot tell you which of its answers were certain.
Where inference stops
Inference runs on models Inferrex trained and serves itself.
When inference does run, it runs on Inferrex's own models: a multi-layer pipeline that classifies fields by business meaning, infers relationships, and proposes cross-system mappings with confidence scores. Your corrections feed back into training, so comprehension compounds rather than resetting with every project.
Trained on specifications, not your data
The models learn from public API specifications. The platform reads structure, not record content, unless you explicitly ask it to.
Read the security commitmentsNo third-party model sees your schema
Inference is served by Inferrex, on Inferrex's own hardware. There is no external provider in the path to route around later.
Runs where you run
In Enterprise and Sovereign deployments inference routes inside your own environment. Nothing leaves it.
Including with no internet at all
Air-gapped deployments use InferrexAI exclusively, with no external calls, enforced at both the network and the application layer.
An inference you cannot audit is an opinion.
Every AI decision is logged and queryable — who, what, when. Classifications carry confidence, mappings are proposed rather than applied silently, and a change the platform wants to make to a live pipeline is presented for approval rather than taken.
That is the difference between a platform that uses AI and a platform you can put in front of an auditor. The comprehension is only useful if the organisation is allowed to act on it, and it is only allowed to act on it if the reasoning is inspectable.
The question was never whether the model is clever.
It is whether you can see what it decided, and whether it ran somewhere you are allowed to run it.
Point it at something ambiguous.
Connect a real provider in the Beta and look at what classifies deterministically, what needed inference, and how confident it was.