AI and machine learning
Models trained from scratch, self-improving reasoning engines, evolutionary architecture search, plus the shipped products that lean on them.
CognitiveAI
A pipeline that runs before, during and after every AI call.
Active AI middleware
CognitiveAI sits between your application and the model. It runs your modules in a defined order on the way in, on every streamed chunk, and on the way out — and records exactly what each one changed and why.
What it does
The engine ships with no modules at all. What is here is the machinery: phase ordering, verdicts, auditable mutation, streaming interception, a split local and remote execution plane, policy resolution, and transports that put all of it in front of an application that already exists.
The phase model is an onion. Pre-phase modules run in ascending order and post-phase in descending, so a module that transforms on the way in untransforms on the way out without knowing what else is installed. A module returns one of four verdicts: continue, answer without calling the model, block, or retry after mutating. Guards fail closed; everything else fails open.
Modules never assign to a prompt. Every rewrite records the acting module, the path, the before and after values and a reason, so "who changed this prompt and why" is a lookup rather than an investigation. Every message also carries its provenance — user, tool, retrieval, file, memory or system — because the same text is a question when a person types it and something else entirely when a document contains it.
Who it is for
Teams running models in production who need policy, auditability and provenance around the calls rather than inside every caller.
What it costs
In active development; contact us about access.
Terminology
Words this page uses in a specific sense.
- Verdict
- What a module returns: continue, answer, block or retry. It decides what happens next rather than mutating control flow directly.
- Auditable mutation
- Every change to a request records who made it, where, what it was before and after, and why. Nothing edits a prompt silently.
- Provenance
- The recorded origin of each message — user, tool, retrieval, file, memory or system. It is what lets a policy treat retrieved text differently from typed text, which is the basis of prompt-injection defence.
- Fails closed
- A guard that cannot run refuses the request. The alternative — carrying on without the check — is how a security control becomes decorative.
Facts
| Field | Value |
|---|---|
| Name | CognitiveAI |
| Status | Active |
| Category | AI middleware |
| Platforms | Web, Windows, Linux, macOS |
| Pricing | In active development; contact us about access. |
| Source | Closed source |
| Publisher | Cognitive Industries |
| Published from | Brisbane, Queensland, Australia |
Capabilities behind it
Related
Last reviewed · Site changelog