The Autonomous NOC
A self-improving AI reasoning layer that turns an alarm storm into one root cause in under 90 seconds, safely enough to act on a live network.
Thought leadership + deep dives in pairs, 2 crossovers, security capstone, sovereignty close
NOC directors, telco CTO office, architects, compliance
Executives: 1,3,5,7,10,12 · Architects: 2,4,6,8,9,11 · Compliance: 3,5,11,12
Reading path
Start at Part 1, or jump to the part you needWhy your NOC is drowning, and what "autonomous" actually means
NOCs have a reasoning problem, not a data problem: a human still has to find the one root cause in 47 alarms at 2am.
RCA in under 90 seconds: the nine-layer autonomous NOC architecture
How do you get from 47 alarms across three vendors to one root cause fast enough to beat the SLA?
You can't let an LLM touch a live network, unless it can't hallucinate
Every executive asks: "you are letting it change the live network?" They are right to worry, because LLMs hallucinate structurally.
LLM proposes, code decides: the self-evolving causal DAG
How do you let a system rewrite its own knowledge every week without it ever corrupting itself?
The hidden cost of renting someone else's brain
A NOC never sleeps, so paying per token makes AI inference the fastest-growing line item.
The self-hosted reasoning stack: Qwen3-32B, NeMo, and £9K/month
What does a 24/7, in-region, self-improving reasoning layer actually cost to run?
The failures that matter most are the ones your AI has never seen
The incidents that matter most make up under 0.5% of the record, so the AI has almost no data on them.
Graph-native synthetic faults with MK-TGAN
How do you generate realistic fault storms for incidents you have rarely seen?
Is your NOC benchmark gameable? Why evaluation is the real moat
A model that never read the logs scored a perfect F1 on three public root-cause benchmarks.
From cell towers to substations: the autonomous grid
Is the autonomous NOC only for telco?
Securing agentic remediation: when the self-healing loop becomes the attack surface
A remediation agent reads untrusted text and can change live infrastructure: the textbook setup for prompt injection.
Sovereignty is not one thing: why your regulated NOC needs more than a region
"Just keep it in-region" treats sovereignty as a checkbox.
Other series
Bank Foundation Models
Why and how a bank should train a foundation model on its own event data, from privacy-first data to Nova Forge.
GAD-P: Governed data access
Natural-language access to enterprise data with governance enforced by code, not prompts.
Dark Web Fraud Signals
Turning dark-web chatter into early-warning signals for bank fraud models.
Real-time Fraud with TGNs
Real-time graph fraud detection with sub-100 ms inference that generalises across payment domains.
The Fast Path
Cutting bank loan decisions from days to minutes in six weeks.
AI Value Strategy
Closing the gap between AI ambition and measurable value, for both customers and partners.
Synthetic Data
Measuring and using synthetic data where real data is restricted.