It tells you when to trust it
How far to trust it and the unknown-data warning come with every answer, not in a separate report. Uncertainty is calibrated where the maths allows, and it says where it doesn’t.
COGNIX · PREDICTIVE AI
Predictions that tell you when to trust them.
A model that is wrong sounds exactly like one that is right. CogniX doesn’t.
REQUEST A DEMO
v3 · Available for pilots
IN SHORT
What is CogniX?
CogniX is a predictive AI engine that turns your data into a prediction service in production: it detects anomalies, predicts values, classifies, clusters and decides. Every answer comes with how far to trust it, a warning if the data is outside what it knows and which variables weighed in. It runs on your network or at the edge, with no telemetry.
THE PROBLEM
What problem does it solve?
- You hear about failures once the line has already stopped.
- You send technicians to check things that were fine.
- You don’t know whether that value is normal for that machine.
- Your alerts get ignored. An alert nobody can check ends up muted.
A bare answer is a dangerous answer: without knowing how far to trust it, you can’t act.
HOW IT WORKS
How does it work?
Give it your data
A CSV or a Parquet file. For anomaly detection, normal data is enough, with no labels.
CogniX publishes the service
It trains and puts it into production without you writing code. First, it checks that the service answers exactly what was measured during training.
Every answer, in four parts
The decision, how far to trust it, a warning if the data is outside what it knows and which variables weighed in. For single readings, batches or a continuous stream.
When your data changes, you retrain it and the new model goes in without stopping the service. Before promoting it, you compare it with the one in production.
CAPABILITIES
What can it do?
| Capability | What for |
|---|---|
| Anomaly detection (DesviantiX) | The unusual, without labelled examples of failures or attacks. It tells the new-but-plausible apart from the anomalous, and follows each device separately. |
| Regression with intervals (RegreX) | A future value and the range it will move within. |
| Calibrated classification (ClassiX) | Which category something belongs to, with probabilities that mean what they say. |
| Clustering (ClusteriX) | What resembles what, to segment equipment, customers or behaviours. |
| Decisions that learn (DecidiX) | Which option to choose, learning from what worked, and a safe option when the data looks like nothing it knows. |
| Rhythm and silences (RitmiX) | Which device has gone quiet or broken its rhythm. In development. |
WHAT SETS IT APART
What makes it different?
It tells you what weighed in
Which variables weighed in each answer: exact where the model allows, and marked as indicative where it doesn’t. No explanation is sold as a cause.
What you measured is what answers
Every training run records its expected answers, and the production service has to reproduce them before it is published.
Your models are yours
In your deployment, each model and its configuration can be downloaded in standard, readable formats such as ONNX.
On your infrastructure
The prediction service runs on your network or at the edge, offline and without sending us anything. It even runs on a Raspberry Pi 5.
Fast where it matters
It answers in milliseconds on the data path, and your AI assistant queries it over MCP to explain what happened. Whatever the console does, the command line and MCP do too.
USE CASES
What is it used for?
| Sector | What it solves |
|---|---|
| Water and gas | Leaks, anomalous consumption and metering fraud. |
| Energy | Anomalous behaviour of installations and forecast values with their margin. |
| Industry | Deviations in the behaviour of machines and processes, with the reason why. |
| Networks and cybersecurity | Anomalous latencies and intrusions without prior signatures. |
| IT operations | Metrics drifting from the normal, with the device and the variables that weighed in. |
EVIDENCE
How do we measure it?
Measured in the lab
Reproducible with the command included in each test.
- Intrusion detection (public NSL-KDD dataset, Sep 2026): trained on normal traffic only, without a single label, and at the same false-positive rate (4%), DesviantiX catches 84.0% of attacks; One-Class SVM, 79.9%, and Isolation Forest, 72.6%. At a 1% false-positive rate, One-Class SVM comes out ahead: 63.8% against 58.2%.
- Performance (Aug 2026, Ryzen 7 7800X3D, 8 cores): up to 53,700 predictions per second for anomaly detection, with a p99 of 2.84 ms, on a light model; the heaviest one measured reaches 19,400 per second. We measure it again on your hardware.
What we don't promise
To always be right. If your «normal» data has anomalies in it, detection drops: with 20% contamination, to just over a third. We measure it and tell you, and we help you clean the baseline before training.
DEPLOYMENT AND PRICE
Where is it deployed and how much does it cost?
- On your network or at the edge, as a standalone product, on an x86 or ARM Linux server.
- Inside NovaeOps, where it puts the AI verdict on the readings. There it is billed per delivered verdict.
- Price: by agreement.
STATUS
What stage is it at?
CogniX is at version 3 and available for pilots. We are looking for one or two reference customers for anomaly detection.
FAQ
What people usually ask us about CogniX
What is unsupervised anomaly detection?
Learning what normal behaviour looks like for your equipment and flagging what falls outside it, without needing examples of past failures or attacks, which almost never exist.
Do I need data scientists?
Not to get started. A domain is created from a CSV and CogniX trains and publishes the service. If you want to go further, we’ll go with you.
Why not use a language model for this?
Because it takes seconds and burns CPU on every reading. CogniX answers in milliseconds on the data path, and the language model sits on top, over MCP, to explain.
What if my normal data is dirty?
Detection gets worse, and we measure it. We help you clean the baseline before training.
Are the models mine?
In your own deployment, yes: each model and its configuration can be downloaded in standard, readable formats such as ONNX. Inside NovaeOps, the model lives in the platform and what you receive are its answers.
Does it work offline?
Yes, in production: the prediction service runs on your network or at the edge without a connection and sends us no telemetry. For training, the first installation downloads its dependencies; if your network is air-gapped, we prepare it with you.
What does it need to run?
An x86 or ARM Linux server (a Raspberry Pi 5 will do) and the ONNX Runtime library. Python is only needed for training. Today it runs as a single instance: if you need high availability, we look at it in the pilot.
REQUEST A DEMO
Got data and alerts nobody looks at?
We’ll show you CogniX with a dataset like yours.
REQUEST A DEMO