ProjectsAgents
Digital Twin AI Ops
A live digital twin of the server estate, where agents know the blast radius before they act and every fix passes deterministic gates and a human.
- Year
- 2026
- Area
- Agents
- Built with
- Apache Flink, Neo4j, ClickHouse, MongoDB, Qdrant, OPA
Telemetry from the server estate is streamed through Apache Flink into a live Neo4j graph of servers and how they connect, with real-time metrics in ClickHouse. Because the topology is always current, the blast radius of a failure is already known before anyone, or any agent, does anything.
How it works
- Sense and normalise. Raw telemetry is mapped onto a canonical model, so every source speaks the same language.
- Reason. Multiple agents work over the twin: one correlates the failure with the topology, another proposes a fix and the evidence for it.
- Act safely. The model can only choose from typed actions, never free-form code. Every action passes seven validation gates, including an OPA policy check and human approval, before it runs.
- Verify and learn. A critic reviews the outcome, failed steps roll back as a saga, and verified results go into semantic memory in Qdrant, so the next incident starts from what worked last time.
Why it matters
Most “AI for operations” demos stop at a suggestion. The hard part is closing the loop: acting on real infrastructure without ever doing something you can’t undo. That is what the gates, the human checkpoint and the rollback are for.