Use Cases / Orion Edge

Use Case: Orion Edge AI-RAN Intelligence

AI-RAN intelligence platform with agentic LLM orchestration, edge telemetry, and autonomous network optimisation across 8 sprints.

Version UCT-3.1-ORION-v1.0 Date 2026-05-03 Owner Uued Viljapuuaiad Status Active Authority Tier 2

Purpose

This document describes the Orion Edge AI-RAN Intelligence Platform — what it does, how it works, what it depends on, and what the critical path to delivery looks like. It follows the Nexus Use Case Template UCT-3.1 structure.

Field Classification

Marker Meaning
[M] MANDATORY — requires confirmed information from client or domain research. Cannot be invented.
[D] DERIVED — UV proposes based on experience. Client confirms or corrects.

Section 01 — Overview [M]

Field Value
Use Case ID UC-OE
Client Reference [M] CodeZero / Orion Edge Internal
System Name [M] Orion Edge
Domain [M] Telecommunications — AI-RAN Intelligence
One-Line Description [M] Multi-sprint AI-driven RAN intelligence platform spanning telemetry baseline through agentic closed-loop orchestration and path intelligence.
Status [M] Production (Sprint 1–8 complete)
Target Delivery [M] Phased delivery — Sprint 1 through Sprint 8 delivered Q4 2025 – Q1 2026

Summary [M]

Orion Edge is an AI-driven RAN intelligence platform built for edge deployment on NVIDIA Nano ZGX GB10-class hardware. It evolves across eight sprints from foundational QoS telemetry baseline through traffic classification, congestion forecasting, policy recommendation, security-resilience fusion, agentic LLM orchestration, SD-RAN intent synthesis, and path intelligence.

The platform ingests real-world 5G telemetry datasets (OTA srsRAN, TRACTOR, TelecomTS), applies ML models for inference, and produces deterministic policy recommendations governed by audit-traced verification loops. Every recommendation passes through deterministic governance before acceptance — the LLM proposes, Orion verifies.

The system is designed as a cohesive, modular platform where each sprint adds capability without rewriting prior work, culminating in an edge-deployable AI-RAN PoC with full observability, security gating, and agentic orchestration.


Section 02 — Actors [M]

Role [M] Description [M] Access Level [D] Frequency [D]
RAN Operator Monitors network health, reviews policy recommendations, acts on alerts Full dashboard + API access Continuous / real-time
Security Analyst Reviews threat-adjusted risk scores, MITRE/FiGHT mappings, security gate decisions Security dashboard + /security/evaluate endpoint On-alert + periodic review
Network Planner Uses what-if simulation and forecasting to plan capacity and resource allocation Policy + forecasting dashboards Weekly / on-demand
Platform Engineer Deploys, maintains, and monitors the Orion Edge stack (API, observability, LLM gateway) Full system access including Grafana, Prometheus, Docker Daily
AI/ML Engineer Trains and evaluates models, reviews model governance artifacts, extends sprint capabilities Source code + model artifacts + training pipelines Per-sprint / as needed

Section 03 — User Stories [M/D]

Actor: RAN Operator

Actor: Security Analyst

Actor: Network Planner

Actor: Platform Engineer

Actor: AI/ML Engineer


Section 04 — System Flows [M/D]

Flow: QoS Inference [M]

Trigger [M]: Client sends RF telemetry payload to /infer/qos.

# Step [M] Actor [M] Input [M] Output [D] Decision [M]
1 Submit telemetry RAN Operator / System rssi, sinr, rsrp, distance_m
2 Validate contract Inference Service Pydantic request model Validation pass/fail Reject if invalid
3 Run QoS model Inference Service Feature vector Throughput prediction + risk class
4 Return response Inference Service JSON with prediction + metadata

Result [M]: QoS prediction returned with risk classification (Low/Medium/High).

Exceptions [D]: Invalid payload → 422 with validation errors. Model load failure → 500 with health check degradation.

SLA [D]: < 200ms p95 latency for single inference.

Volume [M]: Continuous — up to hundreds of requests/minute per cell site.

Flow: Policy Recommendation [M]

Trigger [M]: Client sends context payload to /policy/recommend.

# Step [M] Actor [M] Input [M] Output [D] Decision [M]
1 Submit context RAN Operator / System QoS metrics + forecast risk + security context
2 Deterministic policy engine Policy Service Feature context Candidate recommendation
3 Security gate evaluation Security Service Threat context + MITRE mapping Gate pass/fail Block if threat level exceeds threshold
4 Return recommendation Policy Service Action + rationale + audit trace

Result [M]: Deterministic policy recommendation with full audit trace.

Exceptions [D]: Security gate block → recommendation rejected with reason. Missing forecast data → degraded recommendation with flag.

SLA [D]: < 500ms for deterministic path; < 5s if LLM agent loop is invoked.

Volume [M]: Event-driven — triggered by threshold crossings or operator request.

Flow: Agentic Recommendation (Sprint 6) [M]

Trigger [M]: Client sends context to /agent/recommend.

# Step [M] Actor [M] Input [M] Output [D] Decision [M]
1 Submit context RAN Operator / System Network state + policy context
2 LLM proposal generation Agent Service Context prompt LLM-generated proposal
3 Deterministic Orion verification Verification Loop LLM proposal Verified/rejected proposal Reject if governance rules violated
4 Return verified recommendation Agent Service Verified action + LLM rationale + audit trace

Result [M]: LLM-proposed, deterministically-verified recommendation with full provenance.

Exceptions [D]: LLM unreachable → fallback to offline/deterministic mode. Verification loop rejection → logged with reason, no action emitted.

SLA [D]: < 10s end-to-end with LLM; < 500ms in offline fallback.

Volume [M]: On-demand — triggered by complex scenarios requiring reasoning.


Section 05 — Data Model [M/D]

Entities

Entity [M] Description [M] Key Fields [M] States [D] Relationships [D]
RF Telemetry Record Single observation of radio conditions rssi, sinr, rsrp, distance_m, timestamp Raw → Processed → Inferred Input to QoS Model
Traffic Record Labeled network traffic observation source_file, traffic_class, feature vector Raw → Classified Input to Traffic Model
Forecast Record Time-series congestion observation timestamp, kpi_value, risk_score Raw → Forecast → Policy-input Feeds Policy Engine
Policy Recommendation Deterministic action suggestion recommendation_id, action, rationale, risk_score, gate_status Proposed → Gated → Accepted/Rejected Linked to Audit Trace
Security Evaluation Threat-adjusted risk assessment threat_id, mitre_mapping, risk_adjustment, gate_decision Evaluated → Gated Gates Policy Recommendation
Agent Proposal LLM-generated action proposal proposal_id, llm_source, verified, fallback_mode Proposed → Verified → Accepted/Rejected Wraps Policy Recommendation
Audit Trace Immutable record of every decision trace_id, timestamp, source, decision, rationale Created (append-only) Links all entities
Path Health Record Synthetic path stability observation path_id, attribution (ran_likely, transport_likely, upstream_likely, mixed_uncertain) Observed → Attributed Input to root-cause analysis

Business Rules [M/D]

ID Rule [M] Trigger [D] On Violation [D]
BR-01 Every recommendation must pass deterministic governance before acceptance Policy recommendation generated Recommendation blocked, audit trace logged
BR-02 Security gate must evaluate MITRE/FiGHT threat context before policy emission Pre-recommendation Recommendation held pending security clearance
BR-03 LLM proposals must be verified by deterministic Orion loop — LLM never acts autonomously Agent recommendation flow Proposal rejected, fallback to deterministic mode
BR-04 All decisions must produce an immutable audit trace Any inference/recommendation/evaluation System error — trace failure escalated
BR-05 SD-RAN intents are concept-only — no live actuation Intent synthesis Intent tagged concept_only, no downstream execution
BR-06 Model artifacts must include governance metadata (MAE, R², split strategy) Model training Model not promoted without metadata

Section 06 — Screens [D]

Screen [D] Actor [M] Purpose [D] Key Elements [D]
Sprint 3 — Forecasting Dashboard RAN Operator, Network Planner Visualize congestion forecasts and risk scores Time-series chart, risk score gauge, anomaly bands
Sprint 4 — Policy Dashboard RAN Operator Review and simulate policy recommendations Recommendation cards, what-if simulator, KPI impact
Sprint 5 — Security Resilience Security Analyst Review threat-adjusted risk and MITRE mappings Risk fusion panel, MITRE mapping table, gate status
Sprint 6 — Agent Dashboard RAN Operator Monitor agentic proposals and verification results LLM proposal log, verification status, fallback indicator
Sprint 7 — SD-RAN Intent Network Planner Explore SD-RAN + SD-WAN + SASE intent synthesis Intent mapping view, concept-only badge, intent graph
Sprint 8 — Path Intelligence Network Planner Root-cause attribution for path degradation Path health chart, attribution breakdown, stability score

All screens are Streamlit pages located in src/dashboard/pages/.


Section 07 — Integrations [M]

System [M] Direction [M] Data [M] Format [D] Frequency [M] Fallback [D]
Prometheus Out Platform + model metrics /metrics endpoint (OpenMetrics) Continuous scrape (15s) Metrics buffer; alert on scrape failure
Grafana Out (via Prometheus) Dashboard visualization JSON dashboard definitions Continuous Static dashboard export
Local LLM Gateway (vLLM) Both LLM prompts + completions OpenAI-compatible REST API On-demand (agent flow) Offline deterministic fallback
NVIDIA NIM (optional) Both LLM inference via hosted NIM endpoint REST API On-demand Fallback to local LLM or offline mode
Streamlit Dashboard In (reads API) All inference + recommendation results REST JSON User-driven Cached last-known state
OTA/TRACTOR/TelecomTS Datasets In Raw telemetry for training CSV / trace files One-time ingestion per sprint Re-ingestion from source

Section 08 — Infrastructure [D]

Layer Choice [D] Reason [D]
Hosting NVIDIA Nano ZGX GB10 (edge) Edge-first design — no cloud dependency for core inference
Runtime Python 3.10+ (3.12 venv), Flask API Lightweight, portable, ARM64-compatible
Database File-based (CSV, JSONL, joblib artifacts) No RDBMS needed — telemetry is processed in-pipeline, traces are append-only JSONL
Frontend Streamlit + Plotly Rapid operator dashboard with interactive visualization
Observability Prometheus + Grafana (Docker Compose) Industry-standard metrics stack, isolated from app runtime
LLM Serving vLLM (local) or NVIDIA NIM (remote) OpenAI-compatible API; local-first with optional cloud scale-out
Container Runtime Docker + Docker Compose Observability stack and LLM gateway isolation

Constraints [M]

Environments [D]

Environment Purpose
Development Local edge device — build, train, test, iterate
Demo / UAT Same device — simulation replays and operator walkthrough
Production (edge) Deployed on Nano ZGX GB10 — continuous inference and monitoring

Section 09 — Critical Path [D]

# Item [D] Depends On Duration [D] Blocker Risk [D]
1 Sprint 1: QoS telemetry baseline + ingestion pipeline 2 weeks Low — dataset publicly available
2 Sprint 2: Traffic classification (TRACTOR dataset) #1 (pipeline foundation) 2 weeks Low — dataset available, pipeline reusable
3 Sprint 3: Forecasting + observability stack #2 (feature pipeline maturity) 2 weeks Low — TelecomTS available
4 Sprint 4: Policy recommendation + what-if simulation #3 (forecast risk scores feed policy) 2 weeks Low — deterministic logic
5 Sprint 5: Security fusion + MITRE gating #4 (policy engine must exist to gate) 2 weeks Low — threat mapping is derived
6 Sprint 6: Agentic LLM orchestration + verification loop #5 (full pipeline feeds agent context) 3 weeks Medium — LLM serving dependency
7 Sprint 7: SD-RAN intent synthesis #6 (agent framework reuse) 2 weeks Low — concept-only
8 Sprint 8: Path intelligence #7 (platform maturity) 2 weeks Low — synthetic telemetry

Parallel Work (off critical path) [D]

Item Can Start After Duration
Grafana dashboard definitions Sprint 3 Ongoing
Streamlit page development Sprint 3 Per-sprint incremental
Local LLM gateway setup Sprint 5 1 week
Documentation and README Sprint 1 Ongoing

Milestones [D]

Milestone Reached When Target Date
M1: Core inference operational QoS + Traffic + Forecast endpoints live with real data Sprint 3 completion
M2: Policy + Security pipeline Deterministic recommendation with security gating Sprint 5 completion
M3: Agentic loop verified LLM proposal + deterministic verification end-to-end Sprint 6 completion
M4: Full platform complete All 8 sprints delivered, all smoke tests passing Sprint 8 completion ✅

Section 10 — What Is Needed [M]

# Item [M] From [M] Blocking [M] Status [M]
1 OTA_dataset_5G_srsRAN dataset CCI NextG Testbed (GitHub) Yes — Sprint 1 ✅ Received
2 TRACTOR 5G traffic dataset GENESYS Lab Yes — Sprint 2 ✅ Received
3 TelecomTS observability dataset Academic source Yes — Sprint 3 ✅ Received
4 Nano ZGX GB10 hardware access CodeZero lab Yes — edge deployment ✅ Received
5 vLLM / LLM serving runtime Open-source (vLLM) Yes — Sprint 6 agent flow ✅ Received
6 NVIDIA NIM endpoint (optional) NVIDIA No — fallback exists ⏳ Optional

Section 11 — Acceptance Criteria [M/D]

Definition of Done [D]

Criterion [D] Verified By [D]
All P1 stories implemented across 8 sprints Sprint smoke tests (make smoke-sprint*)
Business rules enforced (BR-01 through BR-06) Contract tests + audit trace verification
All inference endpoints operational pytest test suite (30+ tests)
Deterministic simulation replays pass make simulate-sprint*
Observability stack operational Prometheus scrape + Grafana dashboard load
Security gate logic verified test_sprint5_* test suite
Agent fallback modes work (offline, local_llm) test_agent_offline_fallback.py, test_agent_recommend_*
Model governance metadata present Metadata JSON in all artifacts/ directories

Success Metrics (30/60/90 days) [M]

Metric [M] Target [M] Measured By [D]
Inference endpoint availability > 99% uptime on edge device Prometheus up metric
QoS model MAE ≤ 15.0 Mbps Model metadata artifact
Forecast Event AUC ≥ 0.90 Model metadata artifact
Policy recommendation determinism 100% reproducible given same input Simulation replay comparison
Security gate enforcement 0 unverified recommendations emitted Audit trace analysis
Agent verification loop 100% LLM proposals pass deterministic check before emission Audit trace analysis
Test suite pass rate 100% on all contract + smoke tests pytest CI output

Section 12 — Risks [D]

# Risk [D] Likelihood [D] Impact [D] Mitigation [D]
R1 QoS model limited by RF-only features (low R²) Confirmed Medium — baseline only Transparent governance; stronger signal in Sprint 2+ datasets
R2 LLM serving resource constraints on edge hardware Medium High — agent flow blocked Offline deterministic fallback; local LLM gateway option
R3 Dataset drift between synthetic and production telemetry Medium Medium — model degradation Concept-phase acknowledgment; retraining pipeline available
R4 SD-RAN / path intelligence remain concept-only Low Low — by design Explicit concept_only tagging; no live actuation
R5 Large raw datasets not version-controlled Medium Medium — reproducibility gap Re-ingestion scripts; Makefile pipeline targets

Open Questions [M]

# Question [M] Impacts [D] Resolution [M]
Q1 When will live RAN actuation be in scope? SD-RAN intent synthesis (Sprint 7) moves from concept to production Pending — roadmap decision
Q2 Will NIM-hosted endpoints replace local LLM gateway? Sprint 6 agent architecture and deployment model Pending — NVIDIA NIM availability and cost
Q3 What production telemetry sources will replace academic datasets? All model retraining and validation Pending — operator partnership

Section 13 — Cooperator Brief [D]

Field Value
Scope [D] Multi-sprint AI-RAN intelligence platform: 8 inference endpoints, operator dashboard, observability stack, agentic orchestration layer, all edge-deployable
Stack [D] Python 3.12, Flask, scikit-learn, Pandas/NumPy, Streamlit, Plotly, Prometheus, Grafana, Docker Compose, vLLM (optional)
Timeline [D] 8 sprints delivered — all complete as of March 2026
Dependencies [D] Three public 5G datasets (OTA, TRACTOR, TelecomTS); Nano ZGX GB10 hardware; Python 3.10+ environment
Interfaces [D] 8 REST endpoints (Flask), 1 metrics endpoint (Prometheus), 6 Streamlit dashboard pages, Docker Compose observability stack
Quality Bar [D] 30+ pytest tests covering contracts, determinism, metrics, security gating, and agent fallback; Makefile-driven CI; model governance metadata for all artifacts
Handover [D] Complete source code (src/), model artifacts (.joblib + metadata JSON), training/simulation scripts, Streamlit dashboard, Grafana dashboards, Docker Compose profiles, comprehensive README

Section 14 — Proposed Unified Architecture (Phase 2)

> Status: Proposed — under review. This section captures the target architecture for the next phase of Orion Edge deployment, integrating Atlas, NeXus, Chameleon, and deltaPrism into a unified sovereign stack.


Layer 1 — Hardware + Infrastructure

Target Deployment: HP Z4R

Spec Value
RAM 256GB DDR5 ECC
CPU 24C / 48T Intel
GPU NVIDIA RTX A6000
Storage Local NVMe scratch
Cooling Immersion cooled in Oleon Qloe dielectric fluid
Enclosure GRC Micro tank

Edge Compute Characteristics:


Layer 2 — Orion Edge Runtime

Orion remains the core AI-RAN and observability intelligence engine.

AI-RAN Intelligence

Capability Description
QoS inference Real-time quality-of-service scoring
Throughput prediction Capacity headroom forecasting
Traffic state classification Multi-class radio state identification
Forecast modelling Short-horizon RAN congestion forecasting
Congestion anticipation Proactive threshold-based alerting

Observability Intelligence

Capability Description
ASN telemetry ingestion Autonomous System Number path telemetry collection
BGP path analysis Route stability and path attribution
RTT and latency analysis Round-trip time baselines and anomaly detection
Path attribution Root-cause correlation to specific BGP transitions
Transport fault modelling End-to-end path failure characterisation
Radio-layer correlation Correlating RAN events with transport-layer observations

Deterministic Governance

Control Description
Policy simulation Before/after policy impact modelling
Weighted scoring Multi-factor recommendation confidence
Rollback logic Revert path on negative outcome
Accept / reject / defer Gated decision states for operator approval
Audit trace generation Every decision logged with full input context

Security Intelligence

Capability Description
Anomaly detection Statistical and ML-based deviation detection
Threat-adjusted risk scoring MITRE FiGHT-informed risk weighting
MITRE FiGHT mapping Tactic and technique attribution for telecom threats
Resilience scoring Network-wide resilience quantification
Security-aware gating Security posture gates in recommendation pipeline

Layer 3 — Atlas Sovereign Intelligence Layer

Atlas becomes the sovereign reasoning and orchestration layer overlaid on Orion.

3.1 Telecom Reasoning

Atlas provides causal reasoning over Orion's model outputs:

3.2 Recommendation Generation

Atlas translates inference results into actionable operator recommendations:

3.3 Deterministic Governance

All Atlas recommendations pass through Orion's existing governance philosophy:

Control Description
Deterministic rules Hard rule checks before output
Threshold checks Confidence and value range gates
Explainability scoring Attribution required before output
Audit logging Every recommendation logged with inputs
Confidence gates Low-confidence outputs flagged, not suppressed
Orion verification Atlas proposals verified by Orion deterministic engine

Atlas must never operate as uncontrolled autonomous AI.

3.4 Executive Narrative Generation

Atlas provides the executive storytelling layer:


Layer 4 — CodeZero Sovereign Stack

NeXus (Orchestration + Anomaly Intelligence + Form Flow Layer)

> Formerly: Wormwood. Now unified under the NeXus brand.

NeXus becomes the orchestration backbone, the deterministic anomaly detection engine, and the operator form flow handler for Orion.

Orchestration Responsibilities:

NeXus Rule Intelligence — Anomaly Detection Rule Domains:

Domain Example Rules
BGP / ASN Route flapping, unexpected ASN transitions, prefix instability, path-length anomalies, suspicious upstream changes
RTT / Latency p95 latency spikes, jitter instability, asymmetric latency patterns, RTT variance escalation
Packet Loss Burst packet loss, sustained loss thresholds, loss correlated to ASN changes, transport instability detection
RAN SINR degradation, RSRP instability, throughput collapse, congestion-risk escalation
Security Threat-adjusted anomaly escalation, MITRE FiGHT correlation, suspicious routing patterns, resilience degradation

Example NeXus Workflows:

*Workflow 1 — ASN Path Instability:*

ASN path instability detected


  -> NeXus rule triggered


  -> Orion performs attribution


  -> Atlas generates reasoning


  -> Operator receives governed recommendation


*Workflow 2 — Congestion Risk Spike:*

Congestion risk spike


  -> Forecast confidence > threshold


  -> NeXus approval chain


  -> Atlas proposal generated


  -> Orion deterministic verification


Chameleon / Grafana (Presentation Layer)

Interim Dashboard: Grafana

Grafana is used as the operator and executive dashboard layer until ChameleonBI is production-ready. This applies to all real-time telemetry views, RAN health monitoring, security posture dashboards, and latency analytics.

Dashboard Tool Notes
AI-RAN health views Grafana Interim
ASN path visualisation Grafana Interim
Latency heatmaps Grafana Interim
Forecast dashboards Grafana Interim
Security posture dashboards Grafana Interim
Board-level reporting Grafana Interim
White-label / multi-tenant Chameleon Target state (pending ChameleonBI)

Target State: Chameleon

Once ChameleonBI is available, Chameleon replaces Grafana as the white-label telecom presentation layer:

deltaPrism (Governance + Lineage Layer)

deltaPrism provides immutable audit and governance for every Orion decision and Atlas proposal.

Audit Trails — per event record:

Model Governance:

Security and Compliance:

Reference: https://codezerogroup.com/orion


Architecture Summary

Layer Component Role
L1 — Hardware HP Z4R + GRC Micro (Qloe immersion) Sovereign edge compute, RTX A6000
L2 — AI Runtime Orion Edge Runtime AI-RAN, observability, governance, security intelligence
L3 — Intelligence Atlas Telecom reasoning, recommendations, governance, narrative
L4 — Orchestration NeXus (Wormwood) Workflow, event routing, anomaly rule engine, form flows
L4 — UI Grafana (interim) → Chameleon Operator + executive dashboards (Grafana until ChameleonBI ready)
L4 — Audit deltaPrism Immutable audit, lineage, model governance, compliance

Classification Summary

Section Classification Notes
01 Overview Mandatory Cannot proceed without this
02 Actors Mandatory Must know who uses it
03 User Stories Mixed P1 mandatory, rest derived
04 System Flows Mixed Triggers/steps mandatory, error handling derived
05 Data Model Mixed Entities mandatory, lifecycle/validation derived
06 Screens Derived UV proposes entirely
07 Integrations Mandatory Must know what connects to what
08 Infrastructure Derived UV proposes unless hard constraints exist
09 Critical Path Derived UV calculates from dependencies
10 What Is Needed Mandatory Blocker list
11 Acceptance Mixed Success metrics mandatory, criteria derived
12 Risks Derived UV identifies from analysis
13 Cooperator Brief Derived Internal UV document

Version History

Version Date Changes
UC-OE-1.1 2026-05-06 Added Section 14 — Proposed Unified Architecture (Phase 2). HP Z4R + RTX A6000 target, Atlas telecom reasoning layer, NeXus anomaly rule engine (BGP/RTT/RAN/Security domains), Chameleon telecom UI, deltaPrism audit. Reference: codezerogroup.com/orion.
UC-OE-1.0 2026-05-03 Initial Orion Edge use case document — all 13 sections completed from project repo analysis

*Uued Viljapuuaiad OU — Nexus Platform Documentation*

*Use Case Document: Orion Edge · UC-OE-1.0 · 2026-05-03*