I write code, run a homelab, and ship side projects across web apps, AI tooling, infrastructure automation, embedded hardware, and tabletop game tech. Most of my work lives in private repos, but the breadth is real.

AI and agents. I've shipped a Slack-native agent that helps a marketing team turn quick prompts into publishable LinkedIn posts and on-brand graphics, and a dual-mode RAG chatbot for an arts and crafts business — a public widget for customers and a password-gated internal mode for staff. I write Model Context Protocol servers (the one for Foundry VTT lets language models query and act on a live tabletop game state). I also run a homegrown multi-agent video production pipeline.

Security and identity. I do incident response for small-to-mid Microsoft 365 tenants that don't have a SOC. m365-admin-tool is the open-source CLI I built for that work — a four-command compromised-user triage workflow (doctor / diagnose / timeline / contain) with structured JSON output, graceful degradation per data section, and an opt-in containment path that's logged for the incident write-up. The repo ships with a full playbook, architecture diagrams, and Graph + Exchange permission reference. The homelab side of this work runs Wazuh SIEM, ntopng for flow analysis, Pi-hole for DNS filtering, and a UniFi controller, all logging into a central indexer.

Tabletop game tech. I maintain Vitas Nova, a homebrew D&D 5e campaign wiki I publish out of Obsidian through Quartz onto Cloudflare Pages. I run a Foundry VTT server for the campaign and built a session-recording and live-transcription pipeline so we can replay or annotate sessions afterward. There's also a 5e character sheet editor I wrote that produces print-ready PDFs.

Dev tooling, open-sourced. Small pieces pulled from my daily AI workflow:

• claude-statusline — compact custom status line for Claude Code.
• claude-agents-kit — seven opinionated subagent definitions for code review, debugging, security audits, architecture review, UX review, database analysis, and orchestration.

Hardware and embedded. A Matter-backed Home Assistant deployment handles physical-world automation across the local network. On the embedded side I write ESP32 firmware (soil sensors, Wi-Fi scanners), design printable parts in OpenSCAD, and occasionally tinker with Hyundai/KIA Gen5W navigation firmware when I want a vehicle to do something the manufacturer didn't intend.

Behind all of it, my Obsidian vault is structured as a shared knowledge base where Claude, Gemini, and Codex each get their own scratchpads, with a memory layer that persists across sessions.

Four bare-metal Ubuntu machines, all meshed over Tailscale:

• a Ryzen 9 9950X / RTX 5090 desktop as the primary workstation,
• a ThinkPad P14s Gen 5 for portable work and game nights,
• a production application server running FastAPI services behind Cloudflare tunnels, backed by Postgres 17 and a local Ollama instance for offline inference,
• a home services box running Pi-hole, UniFi, ntopng, Wazuh, Home Assistant, and a Matter server.

systemd-native, no Docker unless there's a real reason to use it.

For API services I'm on Python 3.13 + FastAPI + uvicorn — async by default, Pydantic for typed request/response, OpenAPI generation comes free. Database is Postgres 17; jsonb and ltree cover most of what people reach for Mongo or graph databases for, and logical replication handles cross-host sync without dragging in a CDC tool. Frontend when I need one is React 19 with Vite — server components have made the data-flow story noticeably cleaner than the old SSR dance.

AI work goes through the Anthropic and OpenAI APIs directly, no LangChain abstraction in the middle. Ollama on a local GPU when latency or cost matters more than capability. I write Model Context Protocol servers when I want to give an agent real tools instead of rebuilding function-calling from scratch each time. Slack-resident agents run on slack-bolt in socket mode — no public ingress required, the bot connects out.

Hosts are bare-metal Ubuntu managed by systemd. Services are units, logs come out of journalctl and stay there, no Docker tax unless something genuinely needs the isolation. Tailscale carries all internal traffic (mesh WireGuard with ACLs, no VLAN config to maintain). Secrets live in Bitwarden Secrets Manager and ship to hosts via a small bws-link script — never .env files committed to repos.

Notes and AI scratchpads live in Obsidian, synced across machines with git + rsync. Each assistant (Claude, Gemini, Codex) gets its own scratchpad plus a shared memory layer that persists across sessions, so handoffs between them are cheap.

Total repos visible to this profile build: 12 public, 37 private.

██████░░░░░░░░░░░░░░ Python — 11 repos

█████░░░░░░░░░░░░░░░ TypeScript — 9 repos

█████░░░░░░░░░░░░░░░ HTML — 9 repos

██░░░░░░░░░░░░░░░░░░ JavaScript — 4 repos

█░░░░░░░░░░░░░░░░░░░ Shell — 1 repos

█░░░░░░░░░░░░░░░░░░░ GDScript — 1 repos

• m365-admin-tool — Terminal-first Microsoft 365 and Entra investigation CLI for compromised-user triage. Doctor / diagnose / timeline / contain — structured JSON output, opt-in containment with audit logging. (last push 2026-06-01, —)
• Gluthoric — Profile (last push 2026-06-01, Python)
• claude-agents-kit — Seven opinionated subagent definitions for Claude Code — code review, debugging, security, architecture, UX, db, and orchestration. (last push 2026-06-01, —)
• claude-statusline — Compact custom status line for Claude Code — model, cost, tokens, context, duration, subagent, worktree. (last push 2026-06-01, Shell)

Open to talking about the work. The fastest path is via the projects above or a GitHub issue on one of the open-source repos.

_{Last rebuilt 2026-06-01 14:06 UTC via a GitHub Action that runs every hour.}