Anomaly detection, quarantine and model operations on the data flowing through the fabric — so bad records are isolated before they reach billing, assurance or analytics, and models keep improving on governed data.
The pressureAI is moving into network, care and fraud decisions — and boards and regulators expect it to be governed.
Why it matters to an operator executiveModels are only as good as the data they run on. CORTEX keeps AI on clean, governed data, quarantines what is not and keeps a person in the loop — the difference between a pilot and an AI-native operator.
A foundation capability: it serves every stage, so later use cases go faster.
The fraud and revenue-assurance use cases come together in the SENTINEL command centre →
We start from the KPI an executive owns, not from the technology — then work back to the decisions, data and agents that move it.
AI use-case and risk inventory; value case
Quarantine policy, model governance, MLOps design
Scoring and quarantine on priority flows; model registry
Data science and operations teams on the review queue
More models and flows; MLOps and AgentOps
Same five phases on every engagement; the pilot (phase 3) runs on your data with your team. How the methodology works →
Operators feed billions of usage, network and billing records into models every day. Without a governed processing layer, errors spread faster than people can catch them.
A malformed CDR or duplicate charge reaches billing, assurance and dashboards before anyone reviews it.
Scores arrive without confidence thresholds, quarantine rules or a human in the loop for edge cases.
Retraining is manual and irregular, so accuracy decays as products, tariffs and fraud patterns change.
CORTEX scores records as they move through the fabric, routes them by policy and keeps model performance visible.
Six agents run the model operations; data science owns thresholds, approvals and retraining sign-off.
Insights, confidence, model performance and quarantine status — the screens a data science and model operations team works from.
Figures load live from the demo platform (AI insights, models, quarantine and system performance). Names and amounts come from the demo data set.
| Model | Version | Accuracy | Precision | Recall | F1 | Inference (ms) | Predictions |
|---|---|---|---|---|---|---|---|
Ten production models run with accuracy around the 0.9 mark, and recall trails precision — so some true anomalies still slip through. Suspected duplicates, blacklist matches and device-fingerprint mismatches are the top reasons records are quarantined, and a meaningful share of quarantined records is rated critical.
One model released under governance — registry entry, drift monitoring and documented human oversight — ready for an AI-regulation review.
CORTEX serves the people who build models and the people who rely on their outputs.
Data leadership
ML engineering
Risk teams
Analytics team
How in-pipeline AI processing compares with the approaches operators commonly use.
| Capability | CORTEX | Static rules | Manual analysis | Stand-alone ML |
|---|---|---|---|---|
| Anomaly detection | ML, per record, in the pipeline | Rules only | Sampling | Batch, outside the pipeline |
| Quarantine | Automatic, policy-driven | Manual flagging | Ad hoc | Not available |
| Human review | Queue with evidence | Email alerts | Spreadsheet | Not built in |
| Continuous learning | Retraining on reviewed outcomes | Static | None | Manual retraining |
| Serving | REST APIs | Limited | None | Custom |
CORTEX sits between the Curate and Consume layers: it scores curated data and publishes optimised data, quarantine queues and insights. About the framework →
Capabilities, not promises: model targets and thresholds are set with your teams during the pilot.
Records outside policy are held back before they reach billing, assurance or reporting.
Every prediction carries a confidence score, a model version and an audit trail.
Reviewed outcomes feed retraining, and performance is tracked per model and version.
Models are served through APIs that other accelerators and agents call.
We usually start with one data flow — for example usage records into billing — and one or two models.
Pick the data flow and decisions to protect; agree thresholds, quarantine rules and reviewers.
Weeks 1-2Deploy anomaly and quarantine agents on the flow, stand up the review queue and model monitoring.
30-45 daysAdd flows and models; run MLOps and AgentOps under our managed service if you choose.
Ongoing