Marlon
Paz
Network, Automation and Observability Architect at Core42 in Abu Dhabi, working on data center interconnect and cloud networking. I learned networks from inside the vendors: a year at Cisco as a blue-badge consulting engineer, then IP/MPLS architecture at Nokia in Portugal for smart-grid and railway networks, with Secure ZTP, digital twins, and mentoring the master's students whose research on autonomous EVPN fabrics was published in two peer-reviewed papers. Today I build the agents myself: multi-agent systems that turn intent into verified network state, with every claim checked on the device, and the telemetry that shows what they cost.
Work
agentic-netops: intent to fabricgithub ↗
State intent in plain language; an AGNTCY multi-agent tier (a LangGraph supervisor with mapper, allocator and deployer agents over A2A/SLIM) decomposes it and submits declarative resources. Kubernetes controllers written in Go reconcile them onto a live SONiC EVPN/VXLAN fabric, repair drift, and release what they claimed when intent is withdrawn. gNMI telemetry closes the loop; nothing is reported as deployed until it is verified on the device.
specstride: from specs to tested codespecstride.ai ↗
An autonomous coding orchestrator that takes a Spec Kit feature phase by phase. Nothing advances until an independent LLM critic approves it against the acceptance criteria and the real code. Stuck phases get a diagnostician and a narrowed retry. An opt-in outer loop learns per-phase budgets from its own telemetry and reverts any change that regresses. Standard library only. mixture-of-loops compiles Spec Kit features into provenance-bound, unattended Specstride pipelines.
agent-observability-stack: tokenomicsgithub ↗
End-to-end token and cost telemetry for a multi-agent fleet. Every LLM call instrumented through LiteLLM and OpenTelemetry, streamed to Prometheus, Loki and Tempo, broken down per agent, model and task in Grafana. Every trace also fans out to Arize Phoenix for span-level inspection of LLM calls. GPU-aware, with Telegram budget alerts.
▶ Watch the walkthrough
agentic-ops-bench: which model, which harnessgithub ↗
A benchmark of models against harnesses on real operations tasks: AIOps, NetDevOps, HPC, inference, RAG and tool loops. Local 30B-class models on one RTX 3090 compared with hosted frontier models.
qmd-memory-stack: agent memorygithub ↗
Local three-tier RAG memory over GGUF embeddings, served to the fleet through an MCP gateway. ~19× query speedup on an Intel Arc iGPU via Vulkan.
Publications
co-authored · peer-reviewedTrajectory
Background
CCIE · CKA · Cisco DevNet Professional · CCNP Data Center · Nokia DCFP · NVIDIA AI Infrastructure and Operations · Isovalent Lab Champion
Networks: IP/MPLS, SR, BGP/EVPN, VXLAN, Cisco IOS-XR and NX-OS, Nokia SR Linux and SR OS, SONiC, containerlab
Platforms: Kubernetes, Cilium/eBPF, Go operators, Python, gNMI, Prometheus, OpenTelemetry, Grafana
Agents: LangGraph, AGNTCY (A2A, SLIM), MCP, LiteLLM, Claude Code