Indie developer and business owner.
I run a multi-machine AI home lab (Macs, NVIDIA DGX Sparks, a Windows Surface) and I build: browser/3D games, on-device AI tooling, agent/MCP infrastructure, and automation for my businesses (a creative community platform and a cleaning company).
- ark-runner-v2: A fully procedural Three.js 6×6 expedition truck and camera preview harness from an AI-assisted 3D rebuild of a browser driving game. This is the truck module and review tooling, not the full game.
- one-shot-video-kit: A model-agnostic skill plus tiered pipelines (code-film, mixed-flow, ai-footage) that let a frontier model attempt a finished video in one invocation, from brief to storyboard, build, self-review and score.
- grok-lab-bridge: An MCP server that gives AI bots audited, denylisted command and file access across the machines in my lab. It isn't a sandbox, and the README says so.
- agent-ops-patterns: Small stdlib-only building blocks for running autonomous agents unattended: a hash-chained audit log, circuit breaker, atomic file bus, leased job ledger, and a single-instance cron guard.
- lab-network: SSH from a cloud VM into a home lab behind CGNAT, through an HTTP CONNECT egress proxy, TLS and Tailscale Funnel. Includes connection watchdogs and two competing redesign proposals, neither of them fully deployed.
- spark-fleet-notes: Field notes from running a 235B open model across three NVIDIA DGX Spark boxes with llama.cpp RPC: configs, measured numbers, and the failures that came first.
- ipad-hid-control: Bluetooth HID control of an iPad Pro with measured pointer transfer functions and closed-loop, screenshot-verified clicking. It's research-grade, and its known failures are listed.
- samsung-control: A playbook, a small LAN tv-hub server, and a sample Tizen app for controlling Samsung TVs (and Android tablets over ADB) on a home network.
- capcut-practice: A deliberate-practice loop for short-form vertical editing. It renders ffmpeg filter-graph recipes over a Photos library and keeps a dated log of what worked and what didn't. It doesn't automate CapCut itself.
- Ship real things that run on real hardware.
- Document what broke, not just what worked.
- No vaporware. If it isn't built, it isn't listed.
- Honest READMEs, including what a project is not.
I'm open to interesting engineering conversations, especially about agent infrastructure and running AI on your own hardware. More at closedladder.studio.