"Continuous Improvement" applied to personal intelligence. Reclaiming the power of AI from the cloud to the home.
Coding is thinking — AI just saves us the typing.
True ownership of intelligence. Your AI should serve you, not a corporation's bottom line or data mining operation.
What happens in your home stays in your home. Complete data isolation with zero third-party access.
Leveraging Apple Silicon's unified memory architecture to run datacenter-class models locally.
AI that isn't just a chatbot, but a system that controls your home, manages your data, and knows your life.
Agent 006 reaches production stability. LLM-driven home automation with real-time device control via Hubitat Maker API — lights, locks, thermostats, garage doors, and HSM security. Intent detection classifies events before acting. TTS confirmation loop announces every safety-critical action. Full voice control through the Kaizen iOS app.
A 397-billion parameter Mixture-of-Experts model running entirely on local silicon — only 17B active per token. Three-model Max stack (max:voice, max:deep, max:think), intelligent conditional context injection, MemCore dashboard, and Ollama v0.17.5. The largest model ever run on a single Mac Studio.
Secure external access to KaizenAI via encrypted Cloudflare Tunnels with Email OTP authentication. No open ports, no VPN — just ai.intrac.net behind a Zero Trust gate. Multi-user LLM access tiers for future collaboration.
Full site rebuild on Grav CMS with Caddy reverse proxy. The domain that started it all in 1998 gets a modern platform worthy of the Kaizen vision.
Full restore to Mac Studio M3 Ultra. Optimization for headless rack server operation. Hubitat C-8 Pro migration with 512 GB unified memory powering the entire stack.
Integration of Qwen3-VL and GLM-4.5 Air vision models. Multimodal capabilities enabling the system to see, interpret, and reason about visual input. Later superseded by the Qwen3.5 MoE stack in March 2026.
Knowledge Graph integration and US Metric enforcement added to search proxy. Whisper STT and Kokoro TTS bring natural voice interaction to the ecosystem.
Long before Kaizen had a name, the foundation was being laid. INTRAC.net launched as a cybersecurity firm and internet services provider — running IRC networks, hardening servers, and building infrastructure when "the cloud" was still just weather.
In 1998, when most people were still figuring out dial-up, Shawn McCalla and Rob Rogers launched INTRAC.NET — a cybersecurity and internet services and hosting. The early years were spent running IRC networks, managing dedicated servers, writing security tools, and operating in the trenches of early internet infrastructure. The work spoke for itself in circles where uptime wasn't optional and security wasn't an afterthought.
That same obsession with systems — understanding how every layer talks to every other layer, from bare metal to application — never went away. It evolved. Home automation came next: wiring Z-Wave switches, integrating alarm panels through Envisalink, building Hubitat rule chains that could arm a house and lock every door from a single voice command. Then came the networking: UniFi gear, 10 GbE backbone, VLANs, Pi-hole DNS, Cloudflare tunnels — a proper enterprise stack running in a closet at home.
When Apple Silicon landed with unified memory architectures that could actually hold massive language models in RAM, the final piece clicked. Project Kaizen was born — not as a product, but as a philosophy: take everything learned from decades of infrastructure work and point it at building a truly personal, truly private AI platform. No cloud dependency. No data harvesting. No subscription to think.
Today, a Mac Studio M3 Ultra with 512 GB of unified memory sits in a rack alongside a Synology NAS, a Mac mini M4, and a Raspberry Pi 5 — all connected over 10 GbE, all serving a single vision: sovereign intelligence that lives where you live.
Sovereign AI means nothing if anyone can walk through the front door. Kaizen is accessible from anywhere — but only if you're invited.
Three encrypted tunnels connect Citadel infrastructure to the outside world without exposing a single port. No port forwarding, no dynamic DNS hacks — outbound-only connections from each host through Cloudflare's edge network.
Every protected route requires Email OTP authentication through Cloudflare Access. No passwords stored, no OAuth apps to maintain. You get a one-time code to your email — if your address isn't on the list, you don't get in.
The UDM-SE firewall has zero inbound port forwards for web services. Cloudflare's edge handles TLS, DDoS mitigation, and bot detection before traffic ever reaches the tunnel. The homelab is invisible to port scanners.
Two-layer access control: Cloudflare Access decides who reaches the site. OpenWebUI decides what they can do.
All models, all agents, admin settings, system prompts, memory management. Email OTP: [email protected]
Conversation with Max and selected models. No admin access, no system prompt editing. Added by email to Access policy.
Read-only or limited model access via time-limited Service Tokens. For demos and evaluations.
Max is the primary interface for the Kaizen ecosystem. Not just a chatbot — Max is a fully realized AI personality designed for natural, fluid conversation.
Inspired by the witty and observant AI from the classic film Flight of the Navigator, Max brings a distinct personality to your daily interactions. He knows he runs on a Mac Studio, understands his local environment, and values your privacy above all else.
A native SwiftUI iOS app that puts the entire Kaizen ecosystem in your pocket. Not a web wrapper — a purpose-built client for Max.
Full-screen voice conversation with Max on iPhone. Apple Speech recognition captures your voice, Max responds through Kokoro TTS with live audio visualization. Click to watch the demo.
Full-screen voice conversation with live transcription. State machine cycles through idle, listening, processing, and speaking — with interruption support. Apple Speech framework for STT, Kokoro/Piper for TTS playback.
Pure SwiftUI audio-reactive animations: Orb (cyan-magenta gradient with glow rings and particles), EQ Bars (12-bar circular pattern), Waveform (nested circular rings), and Fluid (organic Fourier-based blob). All respond to live audio levels.
Vector-based semantic memory powered by Mem0 + ChromaDB. The app searches relevant memories before every response and stores new context after each exchange. Browse, search, and manage memories directly from the app.
Pull, select, and delete local models with real-time progress tracking. Image attachments for vision models (up to 4 photos), document attachments (PDF, text), and conversation persistence with search and date grouping.
Connects to 6 backend services: Ollama :11434 · WebSearch :11435 · TTS :8003 · Memory :8100 · OpenWebUI :8080 · Orchestrator :11440
Not renders. Not mockups. Live infrastructure running right now.
A live conversation with Max through OpenWebUI, accessed externally via Cloudflare Tunnel at ai.intrac.net. Max introduces himself — sovereign, local, and deeply connected. This is the interface behind the Zero Trust gate.
14 Kaizen services with 13 running live — Ollama, WebSearch Proxy, Orchestrator, Claude Proxy, Codex Proxy, Z.AI Proxy, Memory, Whisper STT, Hybrid TTS, OpenWebUI, HomeAuto, Caddy, Cloudflare Tunnel, and ComfyUI. All managed from a single dashboard.
The full Citadel homelab dashboard — 4 servers, 3 Cloudflare Tunnels (all healthy), 22 Docker containers, Pi-hole DNS handling 79K+ queries, dual UPS monitoring, Hubitat smart home, and real-time site monitoring across intrac.net and raybridgesjr.com.
Shawn has spent nearly three decades on the bleeding edge of internet infrastructure, cybersecurity, and systems architecture. From running IRC networks and hardening servers in the late '90s to building a sovereign AI platform on Apple Silicon today — the thread has always been the same: own your stack, control your data, build things that actually work.
Project Kaizen is the culmination of that journey — decades of home automation, networking, and relentless tinkering distilled into a single ecosystem where AI serves the person, not the platform.
"We don't need the cloud to be smart. We just need better tools at home."
Expansion to Mac Mini M4 nodes for distributed inference across the homelab.
Advanced generative audio synthesis agent for music production and sound design.
Specialized social dynamics agent.
Autonomous network security monitoring and intrusion detection.
Tiered access to Max and local models via Cloudflare Zero Trust. Trusted users chat, guests demo, owner controls everything.
Self-healing services, proactive maintenance, zero-touch operations, and autonomous agent coordination.