A living model of the network estate — not a static inventory.
Agentic NHI ingests three weekly CSP exports — PM, CM and Inventory — and produces a Network Health Index for every element. Elements are clustered into six behavioural archetypes that evolve week-on-week. An agentic reasoning layer answers questions in the language each team uses, with sources and confidence attached.
Layer 1 — The signal model (NHI)
Network Health Index
A 0–100 weighted composite per network element. Four components, calibrated quarterly.
performance_score
35%
PM KPI composite (avail, RRC, DL util, HO, CQI)
config_conformance
30%
CM fingerprint vs golden baseline
lifecycle_risk
20%
HW age, SW currency, EoL proximity
alarm_burden
15%
Weighted active alarm count by severity
"Weights calibrated quarterly against field incident rate. No vendor-specific tuning needed — the normalisation layer handles vendor differences."
Layer 2 — The element archetypes
Six behavioural archetypes
🟢High Performer(86)
🟡Capacity Ceiling(52)
💡Config Drifter(67)
🔴Aging Macro(44)
🟡Post-Upgrade Settling(61)
🔴Persistent Degrader(38)
"What makes it a living estate — not segments: archetypes track migration (elements moving between personas week-on-week), vendor composition, and the whole estate evolves as the network evolves."
Layer 3 — The agentic reasoning layer
Question, answered with sources
Network Eng
Which Nokia cells are drifting from their config baseline?
847 cells in the North cluster. Primary drift: srPeriodicity non-default on 23% of post-upgrade Nokia BBUs. NHI declining 4pts/week on affected cohort.
NOC
Are any Huawei elements approaching hardware EoL in Q3?
214 Huawei BTS units, install dates 2015–2016, HW EoL flagged. 67% are currently Stable archetype — risk is lifecycle, not active performance.
Capacity Planning
Which high-traffic cells will breach utilisation threshold next month if growth continues?
38 Ericsson eNBs in London South, currently 81–84% utilisation. At current traffic growth rate, threshold breach within 3–4 weeks. Model confidence: 74%.
NHI calculation — from raw weekly fields to a scored network element
Three standard CSP exports. No new infrastructure.
Source
Fields
Cadence
PM weekly summary
8 normalised KPIs per NE
Weekly
CM weekly snapshot
SW version, param fingerprint, drift score
Weekly
Inventory static
Vendor, model, age, EoL status, site class
Monthly
Why "estate" not "inventory"
Static inventory
NEs as isolated rows. No concept of behaviour, clusters, or drift over time.
Living estate
Archetypes have migration paths. A Config Drifter that degrades becomes a Persistent Degrader — that transition is a fleet event, not a database update.
Honest limitations
The model is weekly batch, not real-time. No live alarm correlation. No cell-level subscriber impact data. What it does — score NHI, detect config drift, track lifecycle risk, surface fleet events — it does with precision and transparency.