AI-powered Telecom Operations Assistant

Network operations centers are overwhelmed by thousands of raw alarms during network events, and operators switch between many tools and data sources to understand an alarm or incident. Finding the right data takes time and domain knowledge, delaying root-cause analysis.

An AI-driven triage pipeline ingests alarms and correlates related events, while LLM-based agents are given a narrow set of tools per domain, such as alarms, incidents, and service impact. Prompts and tool routing keep each agent focused, and structured outputs keep responses predictable enough for downstream automation such as work-order creation.

Backend and AI agent engineering: prompt design, tool routing, and service integration. This was professional work for an employer; the system is not my personal property.

• Java
• Spring Boot
• Spring AI
• LLMs
• Tool calling
• Structured outputs
• MySQL

• Separate agents per operational domain instead of one general agent.
• Constrain tools and request parameters so the model cannot call anything outside its scope.
• Prefer structured outputs with strict JSON schema validation so downstream code does not parse free text.
• Use asynchronous event ingestion so peak alarm volume does not block core monitoring threads.

• Requests that could belong to more than one domain.

  Explicit scope and disambiguation rules in prompts and routing.

• Alarm storms producing bursts of correlated noise.

  Deterministic rule-based filtering and deduplication before any probabilistic AI analysis.

• Narrow tool scopes and clear prompts improve reliability more than longer prompts.
• Deterministic rule-based filtering must always precede probabilistic AI analysis in mission-critical telecom environments.