Two-day marathon that touched every layer of the stack: integrated Voxtral TTS, deployed JefeCode code completion API, switched to Let's Encrypt for trusted TLS across 20+ services, rebuilt ComfyUI for Blackwell with video support, deployed six Prometheus exporters, built a JefeHQ service map, migrated services to Titan, launched the Slop Shop AI Comedy Lab with 12 chatbot personalities, hardened Jefebot, and implemented Craig voice capture for the D&D pipeline.

Voxtral TTS and vLLM on Blackwell

Mistral's Voxtral TTS (4B params, 20 preset voices, 24kHz) was integrated as a third TTS backend in JefeAI, AIMemeLord, and GygaxBot. The service calls vLLM's OpenAI-compatible /v1/audio/speech endpoint. Getting vLLM-omni running required building a custom Docker image — the official v0.18.0 image isn't published yet. Also deployed a second vLLM instance (JefeCode) serving Qwen2.5-Coder-32B-Instruct-AWQ for IDE code completion — the first step toward the Inference API product.

Titan Migration Wave

VibeBot4000 and JefeHQ moved to Titan. VibeBot needed a Dockerfile fix, CSP fix, and manual data migration. JefeHQ required extensive env var plumbing (JEFEAI_HOST, FORGEJO_HOST, GPU SSH key, BACKLOG_PATH) and a stale Docker cache that kept serving old code. Also built a git-scanner micro-service on Windows so JefeHQ on Titan can still scan local repos. AIMemeLord pipeline got Ollama + ComfyUI connected via host networking after debugging a Docker bridge issue where Python's httpx couldn't reach the host despite curl working.

jefe.works Expansion

Launched the Slop Shop AI Comedy Lab as a reskinned character generator with JefeAI integration and chat-style episode rendering. Added four new chatbot personalities (Clippy, Shakespeare, Gandalf, Robot) for 12 total. Built a Cloudflare Worker at status.jefe.works for dynamic Twitch live status. Cross-site icon links between jefeworks.com and jefe.works. Hardware details scrubbed from all public sites for physical security. jefehz.org got a modern facelift (dark mode, Space Grotesk, ambient orbs).

Let's Encrypt Replaces JefeCA for Browser Trust

The private CA (JefeCA/step-ca) issued valid certs but every device needed the root cert manually installed. We switched Traefik's ACME resolver to Let's Encrypt with Cloudflare DNS-01 challenge — now internal service domains serve universally trusted certs with zero client-side setup. The tricky part was that Traefik's LEGO_CA_CERTIFICATES env var (set for JefeCA trust) overrode the system CA bundle, so Traefik couldn't verify Let's Encrypt's own HTTPS endpoint. Fixed by bundling system CAs alongside the JefeCA root. Traefik Phases B-D also completed: Forgejo, Grafana, Prometheus, AlertManager, Enclave, JefeAI, and JefeHome all got Docker-label or file-provider Traefik routes.

JefeCode: Self-Hosted Code Completion

Deployed a second vLLM instance serving Qwen2.5-Coder-32B-Instruct-AWQ for IDE code completion. This is the first step toward the JefeWorks Inference API product. The model uses AWQ 4-bit quantization (~20GB VRAM) and coexists with Voxtral TTS on the same GPU via memory partitioning. Planning docs cover Continue.dev IDE integration and Cloudflare tunnel access for external beta testers with service token auth.

Full Prometheus Monitoring Stack

Deployed node-exporter, cadvisor, two postgres-exporters (FreeChat + Enclave), and two redis-exporters. AlertManager was already running but flooding Discord with false positives because Prometheus couldn't reach host services from Docker. Root cause: missing extra_hosts: host.docker.internal in the Prometheus compose, plus UFW blocking cross-network traffic. After fixing both, active alerts dropped from 18 to 2. Also tuned the ContainerHighMemory threshold to exclude GPU containers and deprecated the n8n scrape target.

ComfyUI Rebuilt for Blackwell + Video

Rebuilt the ComfyUI container from ghcr.io/ai-dock/comfyui:latest (v0.2.2, PyTorch 2.4.1+cu121) to a custom comfyui-blackwell:2.0.0 (v0.18.1, PyTorch 2.11+cu128). ComfyUI-GGUF and ComfyUI-Manager custom nodes preserved. Wan 2.2 TI2V-5B model downloading for video generation. Volume mount paths needed fixing — the new image installs to /root/ComfyUI/ while the old one used /opt/ComfyUI/.

JefeHQ Service Map and Titan Migration

Built a new Service Map tab in JefeHQ with an SVG force-directed graph showing all 32 services, dependencies, health status, and clickable dashboard links using internal Traefik domains. JefeHQ itself migrated to Titan — required extensive env var plumbing and a stale Docker image cache that kept serving old code despite rebuilds. Also built a git-scanner micro-service on Windows so JefeHQ on Titan can still scan local repos.

GygaxBot Voice Capture Pipeline

Implemented Phase 1 of the Craig/Whisper voice capture stage for the JefeAdventure D&D pipeline. Three new commands: !session capture <craig-url> downloads multi-track FLAC recordings, !session speakers shows the Discord-to-character mapping, and !session speakermap configures it. The JefeAI transcription endpoint (faster-whisper on Titan) is stubbed for Phase 2.

Jefebot Resilience and Monitoring

Jefebot's EventSub WebSocket had been silently dead for 20+ hours. Fixed with a 5-minute watchdog, IRC fallback after 10 minutes of downtime, and a message dedup cache to prevent double responses during transport transitions. AlertManager was deployed with a Discord bridge for the #jefesonly channel. Prometheus exporters (node-exporter, cadvisor, postgres, redis) filled the monitoring gaps. Active alerts dropped from 18 to 2 after fixing Docker networking issues and tuning alert thresholds.

What's Next

• Wan 2.2 video generation end-to-end testing once model download completes
• JefeCode Cloudflare tunnel for beta tester access
• OBS headless deployment — stream key needed to go live
• Credential rotation P1s via JefeVault
• NVIDIA Inception submission
• Phase 2 voice capture: faster-whisper transcription endpoint on Titan