Modern organizations generate vast amounts of operational data through IT ticketing systems, including incident reports, service requests, and support escalations. These records often contain valuable signals about recurring issues, system weaknesses, and team performance. However, extracting meaningful insights is difficult because most ticketing platforms are built for workflow management, not analysis. Structured fields are inconsistent, free-text descriptions are unstructured and noisy, and relationships between tickets are rarely captured or queryable. When leadership asks for answers about performance, root causes, or improvement areas, teams are often left relying on manual exports, fragile queries, or spreadsheets—processes that are time-consuming, error-prone, and slow to deliver. To address this, NVIDIA’s IT team built ITelligence, an internal AI agent powered by NVIDIA Nemotron open models and integrated with a graph database. The agent serves two core purposes: first, to uncover hidden patterns in unstructured ticket data using advanced LLM reasoning; and second, to analyze complex relationships across tickets, users, devices, and teams through scalable graph-based queries. This architecture is not limited to IT operations. It can be adapted to any domain relying on ticketing systems—such as security incident response, customer support, or facilities management—where unstructured records must be transformed into structured, actionable intelligence. The system is built on a modular, scalable data pipeline with five key stages: Data Ingestion and Graph Modeling Scheduled ETL jobs pull data from ITSM platforms, endpoint inventories, and identity systems. Instead of real-time streaming, a batch-based approach is used, which works well given the tolerance for eventual consistency. Each entity—like User, Incident, Device, or ServiceRequest—is modeled as a node, and relationships such as OPENED_BY, ASSIGNED_TO, or HAS_DEVICE are defined as edges. This graph structure enables complex, multi-hop queries that would be impractical in traditional databases. Contextual Enrichment Enrichment jobs add context to tickets by joining attributes like user role, department, device type, and access level at the time the ticket was created. This allows for deeper analysis without relying on inconsistent or missing user input. Root Cause Analysis (RCA) Standard ITSM categorizations often fail to capture true root causes. To fix this, the system uses an LLM pipeline to analyze each ticket’s free-text description. The model is prompted to extract a concise list of root cause keywords (e.g., YubiKey, passkey, Microsoft Authenticator) based on the content. The most accurate results were achieved using the llama-3_3-70b-instruct model via NVIDIA NIM. These AI-generated RCAs are stored as node properties, enabling precise grouping and analysis. Insight Generation Scheduled jobs use LLMs to generate high-level insights from the enriched data. Prompts are engineered for different use cases: MTTR insights: Identify slow-resolution tickets and summarize reasons for delays. CSAT insights: Analyze low satisfaction feedback to highlight unmet expectations. RCA insights: Surface common patterns across frequently occurring root causes. New hire insights: Uncover onboarding challenges by reviewing tickets from new employees. These insights are linked to organizational context—like team, manager, or service group—so they can be delivered to the right leaders with targeted recommendations. Distributed Alerting and Automated Delivery A rules-based alerting system monitors KPIs such as MTTR, RCA frequency, or CSAT trends. When thresholds are breached, automated notifications are sent to relevant managers with context, affected tickets, and suggested actions. The system can also generate regular AI-powered newsletters, customized per org or leader, summarizing key issues, risks, and opportunities. The user interface is designed for both power and simplicity. While chatbots using RAG are popular, they struggle with complex, multi-entity graph schemas. Natural language queries like “What are the most common issues with VPNs?” can have multiple interpretations—device, user, service, or access type. Translating these into accurate graph queries is unreliable and frustrating. Instead, the solution uses interactive dashboards built with Grafana, powered by the graph database. Static data—metrics, KPIs, insights—is pulled in real time. To automate summarization, a custom API service integrates with the dashboard. When users apply filters (e.g., RCA = driver, Assignment Group = X), the API retrieves matching tickets, sends them to the NVIDIA NIM API via a structured prompt, and returns an AI-generated summary. This summary includes key symptoms, recurring fixes, and strategic recommendations—delivered instantly within the dashboard. This approach eliminates manual ticket triage and delivers actionable intelligence directly where teams work. By combining graph modeling, AI-driven analysis, and intuitive visualization, ITelligence transforms operational noise into clear, timely insights. It empowers IT teams to move from reactive firefighting to proactive improvement—using AI not just to analyze data, but to guide decisions at scale. For those interested in building similar agents, NVIDIA Nemotron offers powerful open models and tools through build.nvidia.com. Explore video tutorials, livestreams, and community resources to get started. Follow NVIDIA AI on LinkedIn, X, Discord, and YouTube for updates and best practices.