Event-driven network automation with AI-assisted troubleshooting

The challenge

The operator ran its network the way most still do. People watched dashboards and opened tickets, then worked the same troubleshooting steps every time the same class of problem came back. It worked, but it didn’t scale. Every new service and every performance blip fell to the same finite team, and the gap between something going wrong and someone noticing stretched across dashboards refreshed and tickets triaged. What they wanted was an operation driven by events, where the network responds to what is actually happening to it rather than waiting for a human to notice and react.

For a service provider, that gap is more than a workload problem. It hits revenue and reputation. Every minute a service stays degraded is a minute a customer feels it, and a team that can only scale by hiring hits a ceiling fast. They needed the network to carry more of its own weight.

What we did

We built an event-driven automation platform, Kubernetes-native from the start. KEDA scales the automation workers on demand in response to real events, so the system does work the moment there is work to do and costs next to nothing when there isn’t.

Then we connected it to the events that actually matter. Splunk searches surface operational signals from across the network. New-service bookings that originate in Salesforce arrive as messages on an Amazon SQS queue and flow straight into provisioning, so a sale becomes a provisioning action without anyone rekeying a thing. The network started to respond to what was happening to it, in something close to real time.

On top of that we built the part the operator was most excited about: LLM-assisted troubleshooting. Using LangChain Deep Agents arranged around a blackboard pattern, the system reasons over incoming signals and shares findings across specialized agents working the same problem, driving toward a diagnosis. It keeps a human in the loop at the decisions that warrant one, so the automation earns trust rather than fear. And the moment a performance issue or a provisioning challenge shows up, the platform triggers automated testing at the affected site, turning detection immediately into diagnosis instead of a ticket in a queue.

Why it mattered

Beyond automation for its own sake, the payoff is concrete. Provisioning that starts from a booking means new revenue turns on without waiting in a human queue. Troubleshooting that starts the moment a problem appears means issues get caught while they are still small. And the team’s scarce expertise goes into supervising an intelligent system and handling the hardest cases, rather than repeating the same first ten steps on every ticket. The operation scales with the network now, not with headcount.

The range it took

This is the kind of system that only works when one team is comfortable with Kubernetes internals and an operator’s Splunk and Salesforce data, and equally at home designing an agentic AI system a NOC will actually trust in production. We architected the event-driven core and engineered the integrations. We designed the human-in-the-loop agent layer and tuned it against the messy reality of a live network. Working across infrastructure, data, and AI at once is what let the pieces add up to something an operator could put its name on.