Projects

Home Assistant sends light commands and assumes success. There's no built-in mechanism to verify that entities actually reached their target state. When you're running scenes or scripts that change dozens of lights simultaneously, the system struggles under the load and commands get silently dropped.

HA Light Controller solves this by sitting between your automations and your lights: it sends the command, verifies the result, and runs iterative retry loops with exponential backoff until every entity matches its desired state.

Brightness, color, and color temperature are all verified within configurable tolerances, and presets let you store complex multi-light configurations as single-tap button entities.

Claude Code is a powerful agentic coding tool, but available tooling tends to be fragmented. Individual plugins from different sources, no shared design language, no consistent principles.

Rather than stitching together tools that were never meant to work together, this marketplace builds everything from a single set of design principles. It covers multi-agent orchestration, Home Assistant integration development, GitHub repository maintenance, release pipelines, Linux system administration (~100 tools across 15 modules), design document authoring, and a live testing framework for plugin development.

All of it is open source. Built for personal use, shared because good tooling shouldn't stay private.

LLMs are powerful but context is expensive. When an AI agent needs to find relevant documentation across a large knowledge base, the naive approach (loading dozens of files and hoping the right information surfaces) burns context budget fast and still misses things.

Markdown-Keeper solves this with a SQLite-backed index for markdown documentation that lets an LLM query directly for what it needs: semantic search, concept lookup, link validation, and chunked document retrieval with token budgeting built in.

Instead of sending an AI to read hundreds of documents, you point it at Markdown-Keeper and it finds exactly what it needs at minimal context cost. Built for developers who treat documentation as a first-class part of their workflow.

Linux administration from inside Claude Code. Rather than wrapping commands behind an MCP server, this plugin gives Claude the domain knowledge to act intelligently: per-service skills that explain config paths, expected ports, health checks, common failure modes, and the gotchas that documentation doesn't mention.

94 skills cover the full range of Linux infrastructure — daemons (nginx, sshd, Postfix, WireGuard), CLI tools (awk/sed, curl, jq, nmap), storage (ZFS, Btrfs, LVM, RAID), containers (Docker, Podman, LXC), and more. Claude loads the relevant skill when you mention a service, then uses its own shell access to act on it.

The /sysadmin command takes a different approach: a guided interview that understands your workload, constraints, and experience level, then recommends a complete server stack with setup order. No build step, no MCP server, no dependencies.

Most dotfile backup solutions are CLI-only, require Git familiarity, or assume a level of Linux expertise that puts them out of reach for intermediate users.

DFBU takes a different approach: a proper graphical desktop application built with PySide6/Qt that makes backing up and restoring configuration files straightforward regardless of your command line comfort level. No Git required, no complex setup.

Mirror and archive modes, backup verification, and profile management give you the flexibility of a serious backup tool. What started as a focused dotfiles utility has expanded into a general-purpose configuration file backup solution for Linux desktops.

Text processing on Linux usually means chaining CLI tools or opening a heavyweight IDE for something that should be simple. TextTools fills the gap: a lightweight PySide6 desktop app with a split-panel interface for encoding conversion, text formatting and cleaning, find/replace, and file management.

Built with an MVVM architecture and dependency injection, the codebase treats a utility app with the same structural discipline as a larger project. Qt Designer .ui files define layouts, ViewModels handle logic via signals/slots, and services are injected at composition root.

The domain doesn't matter; the standards do. PinkBox is a full-stack Django web application built for my wife to catalog her Barbie collection, and it was built the same way I build everything else: with a proper data model, 704 tests, and AI at the core.

Users add dolls via barcode scan or manual entry, and the app uses the Claude API and web search to automatically identify the doll, classify it against collector tier definitions, and estimate market value. Collection browsing, management, and PDF report generation round out the feature set.

Seven implementation phases, all complete, deployed at pinkbox.l3digital.net. The subject is lighthearted. The engineering isn't.