HPE OpsNexus
Enterprise Operations Intelligence Mesh
🏆 HPE FOCUS AI Ideathon 2026 – Winner (Ranked #1)
A reasoning layer above HPE's existing operational systems that predicts the downstream business impact of a change before anyone commits to it.
A six-layer intelligence mesh that federates 15 enterprise systems through their published APIs, builds an enterprise knowledge graph plus a persistent operational memory, reasons across domains with a LangGraph agent layer over Azure AI Search, and delivers explainable, confidence-scored recommendations into Teams rather than a new portal.
Engineering Brief
- My contribution
- System design, the knowledge and memory graph architecture, agent orchestration, and the Phase-1 pilot plan.
- Engineering difficulty
- Reason across operational domains without replacing source systems or bypassing their access boundaries.
- Outcome
- Research proposal with a scoped pilot. Decision-speed and analysis-effort improvements are targets, not measured results.
Engineering questions
The Problem
Every function runs its own tooling, dashboards and AI, so teams spend an estimated 60–70% of analytical effort assembling context before deciding anything. A major GPU procurement can move through six functions with every dashboard green and still land hardware before capacity is ready. The failure is not execution; it is that nothing reasons across the domain boundaries.
System Architecture
Architecture design
Illustrative paths, not live telemetry.
Architecture overview. No node selected.
Full network reference
Enterprise Systems
ServiceNow · OpsRamp · SAP
Kind: external
Published enterprise APIs are the proposed input boundary. The wider integration design is separate from the narrow Phase-1 pilot.
Incoming (0)
None in this diagram
Outgoing (1)
Telemetry & Events
Logs · metrics · changes
Kind: ops
Operational logs, metrics and changes provide context for cross-domain reasoning.
Incoming (0)
None in this diagram
Outgoing (1)
Connectivity Layer
Normalize · RBAC at ingest
Kind: api
Normalize incoming records and attach entitlement metadata before graph and retrieval ingestion.
Incoming (2)
Outgoing (2)
Agent Orchestration
LangGraph · MCP tools
Kind: agent
Coordinate the planner, registered MCP tools and durable reasoning steps, with human approval as an explicit gate.
Incoming (2)
- Connectivity Layerflow
- Knowledge + Memorydata / reason
Outgoing (1)
- Decision Intelligenceflow / predict
Knowledge + Memory
Neo4j · dependencies
Kind: data
Connect assets, services and dependencies with a memory of prior decisions, fixes and outcomes.
Incoming (2)
- Connectivity Layerdata
- Learning Loopfeedback
Outgoing (2)
- Agent Orchestrationdata / reason
- Decision Intelligencedata
Decision Intelligence
Impact · confidence
Kind: service
Combine graph reasoning and retrieved context into an impact prediction with evidence and a confidence score.
Incoming (2)
- Agent Orchestrationflow / predict
- Knowledge + Memorydata
Outgoing (1)
- Teams / Cockpitflow
Teams / Cockpit
Explainable actions
Kind: client
Surface explainable recommendations in Teams rather than requiring another operational portal.
Incoming (1)
Outgoing (1)
- Learning Looptelemetry
Learning Loop
Outcomes · overrides
Kind: ml
Feed outcomes and overrides back into operational memory. This is a design loop, not live telemetry from HPE.
Incoming (1)
- Teams / Cockpittelemetry
Outgoing (1)
- Knowledge + Memoryfeedback
What I Engineered
- Architected a six-layer mesh that federates 15 enterprise systems (ServiceNow, OpsRamp, GreenLake, Azure, Kubernetes and SAP among them) through their published APIs, with zero changes to any system of record.
- Designed the organizational intelligence layer: an enterprise knowledge graph of assets, services and dependencies, paired with an operational memory graph of decisions, actions, fixes and outcomes.
- Specified a LangGraph orchestration layer with a planner, an MCP tool registry, durable multi-step workflows and an explicit human-in-the-loop approval gate as an architectural component, not a UI courtesy.
- Built the decision-intelligence layer on graph reasoning plus Azure AI Search, returning predicted business, SLA and cost impact alongside the reasoning path, the linked evidence and a calibrated confidence score.
- Made calibration a release gate rather than a report: predictions must hold an expected calibration error of 0.10 or better, so a stated confidence means what it says.
- Scoped a deliberately narrow Phase-1 pilot: change-impact reasoning over ServiceNow + CMDB inside one organisation, with defined KPI gates and pivot/kill criteria before any expansion.
Impact & Scale
- Ranked #1 of all submissions in HPE's FOCUS AI Ideathon 2026
- Advanced to final review before the FOCUS Group leadership panel
- Phase-1 pilot scoped to two systems and one workflow: no new data platform, no migration
Targets, not yet measured
- 40% faster decisions and 60% less manual analysis, gated over an 8-week pilot
- Aimed at the ~65% of engineering effort HPE currently spends assembling context
Evidence & Scope
Architecture, engineering decisions, and the Phase-1 pilot plan are documented in this case study.
Source and internal telemetry are HPE-internal. Architecture, trade-offs and the pilot measurement plan walked through on request.
Engineering Decisions
- 01
Consumed the existing systems of record instead of building a new data platform.
Reading GreenLake, OpsRamp, ServiceNow and SAP through their published APIs avoids replacing the systems of record. Connector development, entitlement mapping and operational support still require engineering work; the design reuses existing platforms rather than introducing a competing data platform.
- 02
Paired vector retrieval with graph traversal rather than shipping a plain RAG assistant.
Vector search finds semantically similar things; graph traversal finds causally connected ones. Vector-only cannot answer what a change affects three hops away, and graph-only cannot read runbooks, tickets and email. Cross-domain operational reasoning needs both, so the hybrid is the technical core of the design.
- 03
Captured entitlement metadata at ingestion instead of filtering only at query time.
If access control is applied late, retrieval can assemble an answer from records the requester is not entitled to see. Binding entitlement metadata at ingestion supports permission-aware retrieval, but enforcement still depends on correct query-time checks, current permissions and tested policies.
- 04
Delivered recommendations into Teams instead of building another dashboard.
Adoption is the real risk on a layer like this, and a portal nobody opens produces no measurable benefit no matter how good the reasoning is. Arriving where people already work is what makes the pilot measurable at all.
Stack
- Neo4j
- Represent assets, dependencies and operational memory as connected records.
- Azure AI Search
- Retrieve unstructured operational context alongside graph relationships.
- LangGraph
- Coordinate multi-step reasoning with an explicit human approval gate.
- MCP
- Give the orchestration layer a defined interface to its registered tools.
Source and internal telemetry are HPE-internal. Architecture, trade-offs and the pilot measurement plan walked through on request.
Related work
Shared engineering patterns across separate systems, not runtime dependencies.
- Hexus – Smart Log Manager
Both place controls around model-assisted reasoning: OpsNexus proposes permission-aware graph and vector retrieval, while Hexus documents guarded RCA and ingestion-time redaction.
Enterprise Internal
- RapidDoc AI
Both make system relationships explicit before synthesis: OpsNexus proposes traversal of an operational graph alongside retrieval; RapidDoc extracts dependencies and call relations from code.
Production Pilot
Open to AI Software Development & Platform Engineering roles
Need someone who ships systems like this into production?