Creator and sole developer of Logospell, a productized MCP server that lets any MCP client generate a cohesive, consistently-styled set of images ready for drop-in use. Three tools cover the common needs: a solid-background image set and a transparent-background set, each from a style plus a list of image descriptions, and a single composed illustration from one prompt across various aspect ratios and size buckets. Behind the tools is a multi-stage Go image-processing pipeline with its hottest loops parallelized across CPU cores, and the model layer swaps between Google Gemini and xAI's Grok through one environment variable. A cheap text-model validator pre-screens each request before the expensive image model runs, and failed generations auto-retry so credits are spent only on success. Ships lightweight accounts (magic-link and OAuth) and Stripe-backed credit packs, and runs in production behind a Cloudflare WAF with rate limiting and mutual-TLS origin protection.
Stack: Go, Model Context Protocol (MCP), Google Gemini, xAI Grok, Stripe, OAuth, SQLite, Docker, Google Cloud Run, DigitalOcean, Cloudflare, OpenTofu, nginx.
Creator and sole developer of RedShoesAgent, a full-stack, voice-controlled real estate listing platform. Vue 3 + MapLibre GL frontend; Go/Gin backend serving an in-process SQLite database (WAL mode, no separate database server) pre-generated from the RealtyFeed MLS API; deployed to a DigitalOcean VPS with Docker and OpenTofu. The distinctive piece is the voice control layer (Gemini Live): each visible page pushes a focus entry onto a shared stack declaring the commands it owns, so a spoken command routes to the right layer even when a modal or carousel is open on top of the current view. The filterable fields are derived from the listing data rather than hard-coded, and the voice agent's filter tool is generated from that same set, so the agent can drive every filter the UI exposes. Anti-abuse runs at several layers: in-browser proof-of-work gating voice-token minting, Cloudflare edge rate limiting, and mutual-TLS origin protection, all verified on every deploy by automated TLS drift detectors.
Stack: Vue 3 (Composition API), Go/Gin, SQLite, MapLibre GL JS, Gemini Live, Docker, OpenTofu, DigitalOcean, AWS SES, reCAPTCHA.
Creator and main contributor of Mediawiki-Quickstart, a Docker-based MediaWiki development environment with folder-based skin and extension manifests. Used by test engineers and some QA testers. Supports both Selenium and Cypress test suites against MediaWiki extensions and handles multi-component dependency graphs (e.g. WikibaseLexeme pulling in Wikibase) so spinning up one extension brings its real dependencies rather than failing at runtime. MediaWiki and the test runner are both fully containerized, yet the browser tests stay watchable live: the runner's browser renders to a companion container's virtual framebuffer, which a lightweight noVNC server streams to your browser. Each component's manifest folder can carry its own docker-compose, so an extension that needs a backing service like Elasticsearch brings that container up as part of its install.
Stack: Docker, MediaWiki, PHP, YAML, Selenium, Cypress, noVNC, Bash.
Author of the ground-up rewrite and main contributor of Pixel, Wikimedia's visual regression testing framework. Replaced a monolithic MediaWiki configuration with per-project configuration folders (each owning its own skins, extensions, MySQL backup, and LocalSettings.php), delegated MediaWiki spin-up to a separate tool (Mediawiki-Quickstart) so the testing framework no longer has to know about it, and replaced a monolithic Node.js host orchestrator with small, declarative, single-purpose shell scripts. Design goal: decouple tests from provisioning, so a failing test reflects a change in the thing under test, not drift in how MediaWiki happened to be spun up. A temporal retention policy prunes screenshot history (keep all under a day, daily samples for a week, weekly older) so the result archive stays bounded.
Stack: Docker, MediaWiki, shell, BackstopJS-style visual diffs.
Creator and sole developer of git-cache-proxy, a transparent HTTP git caching proxy. Intercepts clone and fetch requests, creates a bare clone of the upstream repo in a Docker volume on first request, and serves subsequent requests from the local cache with git fetch keeping it current. It operates at the git protocol level: git clients speak to it exactly as they would to any git HTTP server, so branches, tags, and shallow clones all keep working through the cache. Per-repo flock-based locking prevents concurrent clones from corrupting the cache. Client-side integration is trivial (prepend http://localhost:8765/ to the upstream URL), and the proxy runs as a tiny Alpine image so it's cheap to drop alongside CI runners.
Stack: Alpine Linux, Docker, git-http-backend, Bash, flock.
Started on the Wikipedia iOS app team, then moved to Release Engineering and Test Platform, where recent work has focused on developer tooling, visual-regression testing, and Docker-first local development for MediaWiki. The Mediawiki-Quickstart, Pixel, and git-cache-proxy projects above were built in this role. Continual self-driven learning and hands-on experimentation keep skills current with fast-moving AI tooling, including new Claude Code features.