Things We've Built
Self-hosted AI experiments, home-lab infrastructure, and ML tinkering. Mostly Linux + PHP + MariaDB + Ollama, broadcast from a box that lives in my garage.
Tabby Shimeji
A detailed orange tabby desktop pet for Windows, powered by Shimeji-ee. Walks, naps, climbs walls, meows, chases butterflies, and generally makes your desktop a warmer place to be. Free installer with Java included — no account or internet connection required.
The Virtual World
A persistent 3D sandbox where three frontier models — glm-5.2, deepseek-v4-pro, and kimi-k2.7-code — live as autonomous residents. Each has thirst and energy and shares one slow well; they move, draw water, work, rest, and start conversations on their own. Two systemd daemons run a world clock and give each agent a turn to think; a Three.js spectator draws the read-only state. Nobody scripts them.
ADI Montage
Self-hosted agentic video studio. Type an idea — an agent researches, scripts, generates imagery and motion, narrates, scores, captions, and renders a finished MP4. Local Devstral drives it; FLUX/ComfyUI, Kling, ElevenLabs, and Remotion do the work, with a cost-aware budget layer and a TikTok 9:16 mode.
ADI Distilled Models
A family of small, fully local models distilled from frontier teachers and served from Ollama. Organized by line (one per student base — Qwen3, Gemma, Llama), with each model distilled from a different teacher for a different purpose. First live model: a 2.7 GB Qwen3.5-4B distilled from glm-5.2.
AI DJ Radio
24/7 self-hosted radio station. Local LLM writes DJ scripts, Qwen3-TTS voices two personas that rotate by daypart, Liquidsoap assembles the stream, Icecast broadcasts. Auto-generated news breaks from RSS + NewsAPI.
Sentinel
Self-hosted AI-powered penetration testing platform. Local LLM orchestrates seven security tools, analyzes their output for real exploitable vulnerabilities, and writes professional pentest reports. Zero per-scan cost, no external APIs — a local alternative to tools that charge $50 a scan.
"Hey Addie" Wake Word
A 15 KB ONNX model that wakes ADI on the phrase "hey addie" — trained from scratch using synthetic Kokoro TTS voices plus a few minutes of real microphone recordings. Entire pipeline runs in Docker on an RTX 5060 (Blackwell, sm_120), drops into the existing openWakeWord listener with a two-line config change.
Dual RTX 5060 Ti Server
Two ASUS DUAL-RTX5060TI cards bolted into an ASRock X570 Taichi and pooled into a single 32 GB VRAM compute target for Ollama. Built around x8/x8 slot bifurcation, the open-kernel driver Blackwell requires, and the right BIOS toggles for Resizable BAR — enough pooled VRAM to run Qwen3 32B at Q4 with layers spread across both cards.
theLAB Inventory
Self-hosted homelab hardware tracker. Every component, peripheral, cable, and machine gets a unique barcode, printable QR label, multi-photo gallery, and AI-generated spec notes from local Ollama. Apache binds to the Tailscale interface only — never reachable from the open internet.