The SpiritBox Dev Journey
The SpiritBox Development Journey records dated decisions, tests, releases, and lessons from building Basic, Pro, and the Ultimate Windows local-AI workstation.
How twenty years of industrial project-controls work became a growing local-first software portfolio. Product decisions, failed experiments, working releases, and lessons from building with AI assistance.
SpiritBox Ultimate: The Nine-Service Workstation
After months of multi-service integration, SpiritBox Ultimate v1.12 became the first Windows early-access package. What started as a standalone prototype now launches a coordinated nine-service Docker workstation.
The default fleet covers the SpiritBox application, local model service, voice, PostgreSQL with pgvector, Neo4j, Redis, n8n, SearXNG, and the sandboxed execution service. ComfyUI connects through the endpoint the user configures.
SpiritBox Pro Launch: Natural Voice & First Organic Sales
Pro integrated Piper TTS into the desktop assistant so AI responses could be spoken through locally installed voices.
Before a formal marketing campaign, SpiritBox Pro sold two copies at $49 each. That was small but useful evidence that buyers would pay once for a focused local voice experience.
SpiritBox Basic: Local Chat Without Docker
Completed the first Windows edition around LM Studio: local chat, default Spirit Guides, local conversation history, and no required SpiritBox cloud account. Basic does not use Docker.
Established the foundational Spirit Guides architecture — specialized system personas with custom prompt structures and localized knowledge isolation.
Taming CUDA: The VRAM Relay™ Invention
Anyone who has tried running a 14B LLM alongside Stable Diffusion or Flux on a single consumer GPU knows the dread of the CUDA Out of Memory crash.
A 14B model consumes ~18GB VRAM; a diffusion pass takes ~14GB. On a 16GB or 24GB card, they collide instantly.
To solve this without forcing users to buy \$10,000 dual-GPU workstations, Ryan engineered VRAM Relay™: a deterministic memory-coordination daemon that unloads the LLM state to system RAM in sub-second time during diffusion generation, flushes CUDA caches, and instantly rehydrates VRAM when the user speaks again. A comprehensive 25-claim patent application was filed to protect this single-GPU orchestration mechanism.
The 2,000-Hour R&D Exploration Sprint
Sustained, obsessive experimentation across every frontier of the open-weights AI explosion. Benchmarked dozens of LLM frontends and fine-tuning frameworks: Unsloth, LM Studio, WAN2GP, Cole Medin's Archon, and Dark Factory.
Built rapid-turnaround prototypes, including the HeartMuLa Local Music Studio (a working neural audio pipeline running on local silicon within 7 days of model release) and the CrossCheck 1.0 desktop GIS spatial workspace.
The Genesis: Stop Renting AI, Own Your Silicon
Monthly software subscriptions were adding up while local AI tools were improving quickly. Ryan began replacing recurring services with tools he could run, inspect, and improve on his own hardware.
The working thesis was simple: use cloud services when they earn their place, but keep the core workspace and local models under the user’s control.
Lessons from 500,000 Lines of Solo Vibe-Coding
What you learn when you build a full multi-container AI workstation alone with AI pair programming.
Hardware-First Determinism
Never architect around theoretical infinite cloud resources. Build for real consumer GPUs (RTX 3060 to 4090). A real single-GPU machine creates a concrete design constraint. It forces memory use, model size, and workflow handoffs to be tested instead of assumed.
Strict Container Boundary Isolation
Keep LLM inference, diffusion nodes, databases, and workflow automations in discrete Docker microservices with well-defined REST/WebSocket APIs. Clear service boundaries make failures easier to isolate, diagnose, and recover from without treating the workstation as one opaque process.
7-Day Adoption Velocity
When a new neural breakthrough (like HeartMuLa or Whisper alignment) lands on GitHub, don't wait for enterprise SDKs. Build a small local prototype, record what actually works, and only then decide whether it belongs in a customer-facing workflow.
Organic Demand Beats Marketing
Two Pro sales without a formal campaign did not prove a market, but they did prove that someone beyond the builder saw enough value to pay. The next job is clearer positioning and repeatable customer evidence.
See SpiritBox in Action
Watch system walkthroughs, road trip devlogs, and feature demonstrations on our YouTube channel.
Open Sauce Maker Road Trip Devlog
Behind-the-scenes road trip clips driving down from Alberta to California for Open Sauce 2026.
Open Sauce Fan Tribute Site
YouTube Channel: @SpiritBox-ca
Watch deep dives on local model configuration, Piper voice models, ComfyUI rituals, and Second Brain graphs.
Visit Official YouTube Channel