Agentic NOC for telecom networks. Eleven autonomous agents ingest, correlate and diagnose cross-vendor alarms, then remediate — from live alarm to closed incident, automatically.
One physical fault can cascade into dozens of correlated alarms across vendor boundaries — engineers triage noise, not root causes.
Cross-vendor correlation is manual today. Diagnosis depends on individual engineer experience, not systematic evidence.
Cambium, Radwin and Aviat each need their own dashboard and workflow — no shared model, no cross-vendor view.
Inventory, performance, and live alarm data from multiple radio vendors.
Every vendor’s data mapped into one shared canonical schema.
Alarms deduplicated, correlated, and enriched with topology context.
Triggered when a correlated cluster crosses a severity or cascade threshold.
Evidence gathered, hypotheses ranked, risk assessed, remediation proposed or executed, outcome verified.
Every step is audited. Remediation executes only within a defined autonomy tier — every plan step, tool call, and decision is written to an audit trail.
Agentic NOC application for telecom, supporting autonomous ops.
Common building blocks — not built from scratch per product.
Predictive degradation detection shares the same model lifecycle tooling.
Sits on YantrAI’s data federation layer.
Containerized, GPU-ready, with automated deployment and platform support built in.
Same platform behind Iktara’s other production AI deployments — common data ingestion, deployment, upgrade, and support tooling.
Alarm stream.
Planner picks the next agent.
Agent invokes tools.
Evidence written to the store.
Confidence checked vs. threshold.
Outcome confirmed, or re-plan.
Autonomous / approve / escalate, then postmortem.
Implemented as a LangGraph state graph with a hard iteration cap and timeout, so an inconclusive incident escalates to a human rather than looping indefinitely. Runs execute inside Celery background tasks, never inline in an HTTP request.
The Orchestrator dispatches one agent at a time and every result returns through Reflect — each agent reads what the last one wrote to the shared incident state.
Owns incident state; decides which sub-agent runs next.
Groups the deduped alarm stream into incident boundaries.
Pulls PM counters and config needed to test a hypothesis.
Ranks root-cause hypotheses against evidence; distinguishes physical, configuration, and propagated causes.
Scores blast radius before any action above Tier 0.
Turns an RCA finding into a concrete action or a field-dispatch handoff.
Confirms a fix held — alarm cleared, PM counters back to baseline.
Serves past-incident precedent for other agents at reasoning time.
Ticketing, ChatOps notification, and approval requests for Tier 2+ actions.
Compiles the closed incident’s details, writes it to the knowledge base.
Every action the platform can take is classified into one of four risk tiers, and every plan step, tool call, and decision is written to an audit log.
Automatic.
Automatic, audited.
Requires human approval.
Recommendation only, no autonomous execution.
Flags config mismatch (frequency, ATPC, modulation) between paired radios, even absent an alarm.
Scans PM trends for slow degradation (antenna drift, aging hardware) before it alarms.
Helps with capacity planning using traffic-forecasting modelling.
Blocks unsafe actions before they ever reach the network.
Correlates network degradation with subscriber experience data to flag at-risk customers before they churn.
A defined cluster of 2,000 nodes, a phased shadow-to-assisted rollout, and a completion report scored against agreed KPIs — built on YantrAI, Iktara’s enterprise AI platform.