Skip to content

The story

Having heard a lot about AI — especially for coding — a few months ago I wanted to see for myself what it could actually do. I needed a project, and my old Nextion screen (mostly weather, still on ESPHome and Météo-France) was starting to feel dated. So I decided to replace it — this time with a much more ambitious home-automation setup, on a far more capable display, driven by AI from end to end.

One thing led to another: I added a voice assistant to the screen, then a local “engine” to handle voice home-automation on-device / on-LAN, and semi-local or cloud paths for open conversation. That engine is still a work in progress (of course 🙂) — I put a lot (too much) into it to experiment and better understand how LLMs work: RAG, scoring, multi-LLM routing, MCP, and more. I share it mostly for information: I use it day-to-day with this screen, but not yet for coding, nor for everything I tried to pack in — some of which works more or less well.

I started with Antigravity, then leaned on various models (DeepSeek, MiniMax, Z.ai). Next I tried Claude, which also did a lot of the work, then Cursor more recently. In short: this is my everyday screen project — I am more its architect than its creator — born from my first steps into the world of AI.

So, why a screen, though?

After five years with the Nextion, I wanted to give my mostly-weather screen a facelift, keeping at least the same core goals:

  • Be a clock.
  • Warn me at a glance if rain is coming within the hour — do I leave 15 minutes early so I don't show up soaked at the office? Do I grab an umbrella?
  • Show the forecast for the next few days, so I have a conversation topic for the rare occasions I decide to be social.
  • Do all that on reasonable power, always on (well, whenever I'm actually in front of it), with total freedom over layout, design and logic — not just the stock Home Assistant dashboard — with all the upsides that come with it... and the downsides.

Then home-automation ambitions crept in:

  • A direct readout of the soil moisture in my plant pots / veggie patch.
  • Turn the TV and PC on without lifting my butt off the chair.
  • Control the three living-room spotlights and the bedroom light.

Then, little by little...

  • Control my AC.
  • Operate the roller shutter — again, without getting up.
  • Add "Ok Nabu" wake-word support, so I don't even have to lean over to grab the screen anymore 🙂
  • Polish the voice assistant: a conversation mode, a choice of LLM (local or not), and home commands that are as fast as possible and actually understand me.
  • Display my work schedule so I can read my shift hours at a glance.
  • Get a network remote for my TV — it can always come in handy. (Funny story: Claude Fable 5 completely blew me away on that one — it built the whole thing in two prompts flat. Naturally, I then let it loose redoing every single popup in the project… at the cost of roughly 50% of my 5-hour usage cap per popup, on the Pro plan. Worth it.)

All of that with these design goals in mind. My very first sessions with Gemini, rewriting the old Nextion code, were honestly humbling — it rethought how the data got sent and cut the codebase to a third of its size. So this time I wanted lightness and much better optimization than anything I'd hand-rolled myself:

  • No images anywhere — as light and optimized for the tablet as it gets.
  • Data pushes on the Home Assistant side as gentle as possible (my HA runs on a Freebox box, so I have to stay lean).
  • A fast boot with no display lag — something that just feels smooth.

I also aimed for a modern-feeling interface (I'm in my fifties — don't expect miracles): no separate pages, just popups, everything reachable via a button, a long-press, or a swipe. Heavy color-coding gives at-a-glance readability from a few meters away, even though the screen is really meant to be read from under a meter for the fine print. And I tried to pack the maximum info/controls onto something that stays reasonably clean — yes, I know, "clean" is the part I'm worst at. End goal: a screen that looks decent to the eye, even though deep down I'm way more about function than form.

Companion backend (optional, work-in-progress): vromvrom-engine — multi-agent orchestrator used for voice routing and conversation.

→ The linked projects: related projects · How AI was used: note on AI