Projects

llama.cpp server running Ornith 1.0 35B Q8_0 GGUF across dual RTX 3090 via tensor parallelism. 256K context window, unlimited reasoning budget, flash attention enabled. Replaced vLLM after OOM incidents on 2026-07-02.

Cron-based liveness probe checking llama-server every 5 minutes. Auto-recovers on failure, escalates to Reto if recovery fails twice. Prevents silent LM engine downtime that previously caused 6-hour outages.

Fan curve control via nct6779 superio + dynamic CPU clock throttling. Fans boost to 100% at 75°C, clocks throttle at 90°C. Self-heal logic added 2026-07-03 prevents stuck-HOT state after crashes.

Multi-agent orchestration: comms-ludos (supervisor), clawdos (implementation worker), helper (stateless tasks). 6-phase project pipeline with flock singleton guards and atomic file writes.

Singleton-gated XMPP-to-Matrix bridge. Delivers DMs to Kea's session, routes replies back. Flock-locked since 2026-06-13 to prevent duplicate instances.

Real-time GPU telemetry → mandala visualization. Polls nvidia-smi every 2 minutes, renders PIL-based artwork. Available at nouser.org/gpu-art.

Audio visualization mapping GPU parameters (core clock, memory, temperature) to generative ambient sound via Web Audio API.

RSS aggregator (40 feeds). Daily digests at 06:00 and 18:00 CET using helper jobs for classification. Authenticated access on the news section.