All work
AI Telecom Platform
Gen-AI RAN Operations Agent
An LLM agent that answers engineers’ questions about live radio-network performance and alarms, finds root causes, and runs scheduled work.
- Period
- Apr 2026 – Jul 2026
- Company
- Nybsys
- Scope
- Root-cause analysis · RAG · MCP tools · UI
Overview
An assistant for network engineers. They ask about RAN performance and alarms in plain language, and the agent investigates. It calls tools over the network’s time-series and analytics data, draws on a RAG knowledge base, and works out the likely root cause. It also runs work on a schedule, not only on request. A second model checks each answer before it reaches the user. I built the agent’s tool servers, rebuilt the data layer beneath them, and turned the web client into a production React app.
Architecture
01 Triggers
- Web app (chat)
- Scheduled tasks
- Reasoning & evidence view
02 Reasoning
- LLM agent
- Root-cause analysis
- Extensible skills
- Answer verification
03 MCP tools
- RAN performance tools
- Alarm tools
- Knowledge retrieval (RAG)
04 Data
- Time-series store
- Analytics store
- Knowledge base
What I did
01 What the agent does
- Root-cause analysis: investigates network problems across performance and alarm data, and explains the likely cause, not just the symptom.
- Scheduled work: runs tasks on a schedule as well as answering questions on demand.
- Knowledge base: grounds its answers in a RAG knowledge base as well as live network data.
- Verified answers: a second model checks each answer before users see it.
02 Agent & MCP tooling
- Built the MCP tool servers that give the agent access to RAN performance and alarm data.
- Made tool calling dependable: input validation, clearer tool descriptions, and errors the model can recover from.
- Added answer verification by a second model, and an extensible skills system so operators can shape the agent’s behaviour.
- Set up an evaluation environment to test agent changes before rollout.
03 Data layer
- Moved performance data to a time-series database and alarms to a column store, both suited to fast analytical queries.
- Migrated the historical data into the new stores, and corrected KPI calculations along the way.
04 Product UI
- Rebuilt the web client in React with a new design, sign-in, and session management.
- Made answers transparent: users can see the model’s reasoning, the evidence behind each answer, and usage.
- Kept long-running agent tasks reliable end to end.