Marlon
Paz

Tokenomics Observability

Making agent spend legible: end-to-end token & cost telemetry for a multi-agent fleet — every LLM call instrumented through LiteLLM and OpenTelemetry, streamed to Prometheus, Loki and Tempo, and surfaced on Grafana dashboards that break down tokens, latency and dollars per agent, per model and per task. Local-GPU aware (Intel iGPU/NPU + RTX 3090 eGPU), with Telegram budget alerts.

▶ Watch the walkthrough

Agentic Observability Stack · 13 views on YouTube

AI Governance Control Plane

A vendor-neutral AI governance platform for continuous model & agent inventory, human oversight, control orchestration, and evidence custody across deployment boundaries. Built on a strict invariant — a UI, event or dashboard is never governance authority: consequential changes require current human authority, separation-of-duties checks, an atomic canonical-state-and-audit commit, and retained evidence. Hexagonal domain core with replaceable ports over a PostgreSQL canonical store, a hash-chained audit ledger, a content-addressed evidence store, an OPA policy decision point, and event-driven projectors over NATS JetStream.

Production Multi-Agent Systems

Designing and operating teams of autonomous AI agents that plan, build, review and act — orchestrated with LangGraph over the AGNTCY Internet-of-Agents stack (A2A protocol, SLIM transport, NATS), multi-model through LiteLLM, and fully instrumented with OpenTelemetry (traces into ClickHouse and Grafana). Python 3.13 / FastAPI services, a React 19 frontend, shipped on Docker & Kubernetes.

Agentic RAG & Agent Memory

Giving those agents a memory: a local, GPU-accelerated Retrieval-Augmented Generation layer where the agent itself decides what to retrieve, re-queries, and grounds its answers in cited sources — a hierarchical, three-tier semantic memory over local GGUF embeddings and vector search, served to the fleet through an MCP gateway. Runs on-prem on an Intel Arc iGPU (~19× query speedup via Vulkan), fail-open by design.

Agentic AI & Autonomous Cloud & Network Infrastructure

Converging two decades of networking and automation with agentic AI across the full stack — from the cloud infrastructure and Kubernetes platforms up through the networks that bind them: building AI agents and the systems that connect, secure and observe them (AIOps), open standards like Agntcy that let agents interoperate across the Internet of Agents, and autonomous, self-driving cloud & network infrastructure that runs and heals itself. At the essential core of the stack is reliable, resilient, event-driven automation and observability — the foundation everything else is built on — all with governance that keeps it accountable. The plan favors depth over speed.