AI Solution Architect & DevOps Engineer — Ladispoli / Rome, Italy
Twenty-six years of enterprise infrastructure — Linux, virtualisation, networking, cloud, CI/CD — now spent on a different problem: making AI agents operational on real systems rather than on slides.
What actually interests me is the point where a language model stops being a demo and becomes a piece of infrastructure: when it has to authenticate, read state that changes under it, fail recoverably, and leave a trail somebody can verify afterwards.
Agentic orchestration on HPC. An agentic layer over FirecREST, the REST gateway CSCS (Swiss National Supercomputing Centre) exposes to its supercomputers: an MCP wrapper over the API, a board where the agent is assigned work like a team member — with an execution log and a human review gate — and a local RAG layer over the documentation, so the agent looks parameters up instead of inventing them.
The goal is that a researcher who isn't an HPC expert can submit, monitor and diagnose jobs in natural language. The whole stack runs offline on a laptop, so it can be tried without asking anyone for resources.
FirecREST Hackathon — Automating Your Workflow with FirecREST CSCS / ETH Zürich · 3 November 2026
| agentic-server-platform | Production-grade autonomous server platform: OmniRoute LLM gateway, Multica workspace orchestration, and Hermes Agent runtime in Docker & Kubernetes. Featuring token governance, semantic context compression (RTK), and Cloudflare Zero Trust ingress. |
| firecrest-agentic-workbench | Agentic layer over CSCS FirecREST: MCP wrapper + Multica board + DocMind RAG. Built for the FirecREST Hackathon (CSCS / ETH Zürich). |
| rodecaster-ndi-hx | Turns a USB webcam on a Raspberry Pi 4 into a real NDI|HX source using the hardware H.264 encoder. Reverse-engineered against the NDI Advanced SDK with a RØDECaster Video as the receiver. |
| framework360-mcp-server | MCP server + CLI over Framework360's 141 REST endpoints: clients, orders, chat, marketing, reports — from Claude Code or the terminal. |
| framework360-skill | The skill that teaches an agent to drive that CLI without guessing commands and parameters. |
- A claim without a source isn't a claim. In my projects an unverified fact is
written
[unverified], with a note of where I looked. That applies to what an LLM says too — especially to that. - A reported failure is worth more than a described success. When a container won't start or a call fails, the value is in the exact command and the verbatim output, not in an account of how it should have gone.
- Work lands in review, not in
main. Even when an agent did it. Especially when an agent did it.
Stack: Linux (openSUSE, RHEL, Debian) · Kubernetes · Terraform · Docker · CI/CD · MCP · RAG (LlamaIndex, Qdrant) · Python · multi-runtime agents
📍 Ladispoli / Rome · 🔗 poplme.co