Skip to content

Related Projects & AI Experiment Context

Vigie — Global Weather & Hazard Ecosystem

A companion ecosystem under active development to gather natural hazard alerts and weather across the world: - Vigie API — Python service aggregating global hazards (earthquakes, tsunamis, volcanoes, storms, floods) and weather data into normalized GeoJSON and an ultra-compact binary/text frame for embedded microcontrollers. - Vigie Web (Live at vigie.auxfilsdesidees.fr) — Public interactive web portal to explore real-time hazards on world maps and local forecasts. - Vigie Display — Standalone ESP32-P4 / Waveshare 7-inch touch screen running ESPHome + LVGL, querying the Vigie API directly with no Home Assistant server needed. - Home Assistant integration (custom_components/vigie) — Feeds sensors, events, and automations into Home Assistant to enrich this Tab5 screen.

vromvrom-engine — Multi-Agent Orchestrator

This screen is the physical front-end of a broader experiment. The companion backend is vromvrom-engine — a Python asyncio multi-agent engine used (among other things) for:

  • Routing voice / chat requests between local LLMs and cloud APIs by task type, budget and latency
  • Fast paths for short Home Assistant commands vs light “Discussion” chat (TTS-friendly answers)
  • Hybrid memory (SQLite + ChromaDB) and RAG for project context
  • MCP tool servers for Home Assistant, memory, and IDE assistants
  • Experimenting with scoring, multi-LLM cascades, specialists (web, calendar…), and self-healing loops

The Tab5 and the engine are two sides of the same experiment: the Tab5 is the push HMI + voice satellite; the engine is the reasoning / routing backbone. Neither is finished. Both are actively changing.

Note: The public engine repo is an anonymized OSS mirror of a private day-to-day tree. Architecture is documented, but the codebase evolves quickly as new AI capabilities are tested. If pieces don’t quite fit together, they were often written by different models at different times — and shared here mainly so the Tab5 voice path is understandable, not as a “finished product.”

Older / related workspace material may still mention ServeurHA; for the public link, prefer vromvrom-engine above.

A note on this project

This is a personal project. The goal was never to build a product — it was to explore what AI tools can actually do when given a real hardware target, real constraints, and real debugging cycles.

Everything in this repository — the ESPHome YAML, the C++ code, the Home Assistant automations, and most of this documentation — was generated by AI: - Antigravity / Gemini (Google) — early orchestration, documentation, code review - DeepSeek, MiniMax, Z.ai — generation and iteration passes - Claude (Anthropic) — architecture, C++, debugging - Cursor — recent work (voice path, HMI polish, docs)

The human in the loop defined what the project should do, tested what came out, said “this is wrong” or “close but not quite,” and kept steering. No code was written by hand — more architect than author.

This is not a claim about AI capability or human obsolescence. It's just what the experiment looked like in practice. Some things worked better than expected. Some things required many iterations. A few things still don’t work correctly. That's the honest version.

If any part of this is useful to you — as a starting point, as a reference, or as a “what not to do” — that's good enough.