L'électron rare
0e57f83e7e
Brainstorm 4/4 of today's session. Decisions: - Architecture: native daemons + LiteLLM proxy (mirrors the existing Tower pattern so client contracts stay stable). - Hardware target verified: M3 Ultra 32-core, 512 GB unified RAM, 7.3 TB disk. Ample headroom even for the 72B Q8 LLM. - TTS: Piper, status quo (3 FR voices: tom-medium for the NPC, siwis + upmc for auxiliaries). .onnx files copied verbatim from Tower. - STT: whisper.cpp large-v3-turbo Q5_K, FR-multilingual. Greenfield — no production STT today. - LLM: dual-tier Qwen 2.5 7B-instruct (NPC dialogue, fast) + 72B-instruct Q8 (hints engine, reasoning). LiteLLM routes by model alias. - Network: home LAN, static IP 192.168.0.150 on the MacStudio. Clients (ESP32, atelier) hardcode the IP. Tailscale stays for remote ops. - Cutover: hard, single weekend. P0+P1 prep during the week, P2+P3+P4 over Saturday-Monday. Rollback drill before declaring final. Phased plan (P0-P4) with explicit acceptance gates, ~5-7 h prep + 3-4 h cutover + 24 h soak. Tower keeps host duties (Suite Numérique, n8n). KXKM-AI keeps the RTX 4090 for fine-tuning, just stops serving inference. 3 brainstorms remaining today: playtest harness, hints engine, voice pipeline ESP-SR.
Zacus — Docs
Ce dossier est conçu pour être publié via GitHub Pages (source : /docs).
- Accueil :
docs/index.md - FAQ :
docs/faq.md - Design :
docs/design.md