dani@geneva:~$ ./field-journal

Five worlds. One thread.

Claude Code tutorials, MCP server guides, Proxmox home lab setup, Home Assistant automations, and prompt engineering for developers.

I’m Daniele Messi, Senior Editorial & Digital Experiences Manager at STMicroelectronics in Geneva. I write guides and publish the code and blueprints behind them, and they get read by very different people: agent builders, marketing leads, home-lab tinkerers, Home Assistant power users, and Italian professionals working out the AI Act. Scroll through their worlds and follow the orange line. It is the part they all share.

The AI & MCP Developer

The switchboard

The AI & MCP Developer

“Can my agent notice a broken tool call and fix it before I wake up?”

Engineers wiring language models into real work. They run private workflows, write MCP servers so an agent can reach a file system, a database or a calendar, and they have learned that a demo agent and a production agent are different animals. They want patterns that hold up against messy inputs and hostile text.

Looks for
  • Self-healing agent loops: retries, fallbacks, checkpoints
  • MCP server design and multi-agent throughput
  • Local LLMs running next to cloud models
  • Prompt injection defense for tools that touch real data
Takes away
An agent that fails loudly, recovers quietly, and never leaks what it reads.
{ "jsonrpc": "2.0", "id": 42,
  "method": "tools/call",
  "params": { "name": "read_invoice",
              "arguments": { "path": "inbox/2026-09.pdf" } } }
One MCP tool call. The interesting work happens around it: validation, retries, and deciding what the model is allowed to see.
Read: MCP multi-agent performance →

The B2B Content & Marketing Strategist

The press room

The B2B Content & Marketing Strategist

“How do we publish more, in more languages, without turning the blog into a content farm?”

Managers running the digital presence of large companies. The calendar is full, the CMS is old, and every article has to pass brand, legal and SEO review. They want automation that keeps an editor in charge: drafts that arrive already researched, keywords mapped to real search intent, and a pipeline they can open up and inspect.

Looks for
  • Editorial automation with a human approval step
  • SEO keyword targeting and structured data
  • Content that answers well in search and in LLM results
  • The Astro Content Engine, an open-source pipeline from brief to published page
Takes away
A publishing pipeline the team owns, can audit, and can run on a schedule.
---
title: "Edge AI on a 5 W budget"
lang: en
keywords: [edge ai, tinyml, microcontroller]
status: review   # a person approves before publish
---
A post’s frontmatter. The status field marks where automation stops and an editor takes over.
Open astro-content-engine →

Fresh from the press room

All writing →

The Home Lab Architect

The rack

The Home Lab Architect

“What does this container cost me in electricity, and does it need the internet at all?”

Tech enthusiasts with a Proxmox box in a closet and opinions about idle power draw. They split services into lightweight LXC containers, run small AI models at the edge instead of renting cloud GPUs, and treat the firewall between their own networks as seriously as the one facing the internet.

Looks for
  • Proxmox setup and hardening scripts
  • LXC containers sized for low-power edge AI
  • Internal firewall and network segmentation rules
  • Energy tracking per host and per service
Takes away
A server that stays quiet, stays local, and shows its electricity bill.
pct create 120 local:vztmpl/debian-12-standard_12.7-1_amd64.tar.zst \
  --hostname llm-gateway --cores 2 --memory 4096 \
  --unprivileged 1 \
  --net0 name=eth0,bridge=vmbr0,firewall=1
An unprivileged container with the Proxmox firewall switched on from the first boot.
Open proxmox-homelab-scripts →

The Smart Home Engineer

The blueprint

The Smart Home Engineer

“Can the house decide when to run the dishwasher, from tomorrow’s prices and today’s sun?”

Home Assistant users who outgrew “turn on the light at sunset”. Their automations read solar forecasts and hourly tariffs, move loads to cheap hours, close a valve when a leak sensor gets wet, and raise an alert when an unknown device joins the network. It all runs locally, so it keeps working when the internet doesn’t.

Looks for
  • Predictive energy control with solar and battery
  • Dynamic tariff automations
  • Water leak detection with automatic shut-off
  • Local network and security monitoring
  • Local camera and alarm integrations (I contribute to eufy-sdk)
Takes away
Blueprints they can import in a minute and understand in five.
blueprint:
  name: Run when power is cheap
  domain: automation
  input:
    price_sensor:
      selector:
        entity: { domain: sensor }
The top of a Home Assistant blueprint. Inputs make one automation reusable across different houses and tariffs.
Open automation blueprints →

The thread

Different rooms. Same wiring.

  • Automate the repeatable

    If a task comes back every week, it gets a script, a schedule and a log. People keep the judgment calls.

  • Own the infrastructure

    Local-first where it matters: your data, your hardware, your keys. The cloud gets used where it earns its place.

  • Apply AI where it pays

    Models are tools. They go where they save hours or catch mistakes, and nowhere just for the demo.

Lately

  • Contributing to eufy-sdk

    Open-source eufy cameras and alarms for Home Assistant: a WebRTC command transport for the HomeBase S1 Pro, the HomeBase alarm as sensors, siren support for the HomeBase 3, and a migration guide from eufy_security.

    ha-eufy-sdk on GitHub →

  • Rebuilt this site

    The five worlds you just scrolled through: one particle field and one thread, drawn on a canvas with no libraries.

    See the work →

  • Publishing every week

    New guides on agents, home labs and automation, in English and Italian.

    Read the latest →