Session 10 was the great migration. Services that lived on the Windows dev machine for months finally moved to Titan — a dedicated GPU inference server. Alongside the migration, we designed two commercial-grade architecture docs, added real-time GPU monitoring to the dashboard, pushed every repo to self-hosted git, and wrote 22 alert rules for the monitoring stack. The local machine is lighter, Titan is heavier, and the ecosystem is meaningfully more production-like.
Services Move to Titan
GygaxBot (the D&D Discord bot) migrated to Titan and reconnected to Discord immediately — campaign data from 5 sessions, character sheets, and reference material all copied over. The AIMemeLord content pipeline is deployed and running idle on Titan, ready for the next stream. JefeAI API migration is in progress with the Docker image building and a Cloudflare tunnel created. On the Windows side, we decommissioned the local Ollama instance, the Cloudflare tunnel daemon, and the Ory authentication stack. Jefebot's image generation commands now route to Titan's ComfyUI, which has an actual GPU behind it instead of a CPU pretending.
Inference Provider Router
Designed a health-check-based routing layer for LLM inference: Titan as the primary provider, local fallback for when the server is under maintenance, and cloud as a last resort. The design doc goes deeper than routing — it includes a commercial tier mapping for JefeWorks (code inference as a service), a distillation rights analysis for every model we considered, and a full landscape assessment of what's available. The takeaway: Apache 2.0 models with no branding requirements exist at competitive quality levels, and they fit in our GPU's VRAM alongside smaller utility models.
Code Model Landscape
Audited the open-source code model landscape specifically for a commercial inference offering. Qwen3-Coder-Next 80B (Apache 2.0, 70.6% on SWE-bench Verified) emerged as the top candidate — it fits in VRAM alongside smaller models and carries full distillation rights. The commercial structure maps to three tiers: Fast (autocomplete via small model), Standard (code generation via medium model), and Premium (architecture and multi-file reasoning via the 80B). All recommended models are Apache 2.0 or MIT-licensed with no restrictions on output ownership.
JefeOS Vault Assessment
Combined the OS engineer and security engineer to assess using JefeOS — our custom x86_64 kernel — as a dedicated password vault appliance. The crypto foundation is solid: AES-256, ChaCha20-Poly1305, TLS 1.3, and SSH are all implemented and tested. The gap is process isolation. A single-process kernel with no memory protection is fine for a serial-console vault accessed over SSH, but it's not a general-purpose secure OS yet. The assessment produced a 3-phase plan from kernel-mode MVP (usable today for a single-user vault) through full user-space isolation with memory protection. Both docs feed into the JefeWorks commercial strategy.
JefeHQ Dashboard: GPU and Git Tabs
The dashboard gained two new tabs. The GPU tab shows real-time utilization, VRAM usage, temperature, and power draw from Titan's enterprise NVIDIA GPU, plus a list of running GPU processes. The JefeGit tab pulls repository status from the self-hosted Forgejo instance — last commit timestamps and freshness indicators for every repo. A combined Git view with GitHub/JefeGit/Both toggle lets you see sync status at a glance. Fixed a bug where the GPU tab's auto-refresh timer was stomping on other views when you navigated away.
Git Hygiene
Pushed all ecosystem repos to self-hosted JefeGit (Forgejo on Titan), giving us a second copy of everything outside GitHub. Merged stale feature branches back to master across JefeAI and Jefebot — no more orphaned work sitting in limbo. Ran a security scan across all repos: clean, no tracked secrets. Cleaned stray config files from tracked git history.
Monitoring Overhaul
Overhauled the Prometheus configuration for Titan: added dcgm-exporter as a scrape target for GPU metrics, fixed mismatched job names that had been silently dropping data. Wrote 22 Titan-specific alert rules organized into 4 groups covering GPU health, service availability, resource exhaustion, and inference pipeline status. On the AI training side, generated 68 D&D rules Q&A examples from the 5e SRD for model distillation — structured training data for the rules-lawyer model that will power JefeAdventure.
What's Next
- Complete JefeAI API migration to Titan (Docker image building, tunnel verification)
- Wire up the inference router for automatic failover between providers
- AlertManager deployment for Prometheus rule notifications
- JefeAdventure Phase 1: API router wrapping existing game logic
- Continue SRD distillation (68 of target 428 examples complete)
- Bank account and Stripe setup for JefeWorks revenue