Things I built and broke. I shipped.

A card generator that reads one public repo's commit history and tells you what you were obsessed with — the 3am streak, the 107-day silence, the return.

Snowflake does all of it. An external access integration pulls the repo from inside the warehouse, fifteen SQL views score six storylines and pick exactly one, and a UDF over AI_COMPLETE narrates that single thread — 20 to 140 commit lines, never the whole history. Thirteen SQL checks run before a card exists, so it fails loudly instead of rendering a lie.

A deduction game where you out-question an Oracle and race a Gemini-played wolf to name the hidden rune first.

A SvelteKit app on Cloud Run where a deterministic engine owns every fact while Gemini interprets Asks, plays the wolf, transcribes speech, and voices both characters — never seeing the secret. Provably winnable every round, keyboard- and screen-reader-complete.

A tool that walks the energy supply chain backward, from any U.S. address to the coal mine feeding its grid.

SvelteKit and FastAPI on Cloud Run trace outlet to power plant to fuel contract to mine, surfacing MSHA accident records, EIA production, and EPA emissions for each. Snowflake Cortex powers natural-language queries; a live tonnage ticker shows extraction in real time.

An immersive memoir about fighting to become an engineer when the world wasn't built for you to.

A framework-free browser experience — vanilla JS coordinating Canvas 2D rendering, PixiJS displacement shaders, and GSAP-timed scene transitions, synced to layered Howler audio cues through every shift in bias, endurance, and becoming.

A unified global aircraft registry assembled from public civil-aviation sources, normalized to one schema and stripped of personal data.

FAA is the schema baseline; other registries extend it only when they publish something structurally new. Country differences become config-driven field mappings instead of one-off parsers, so one engine covers every country. Each build ships as an immutable, versioned static snapshot — no database to run or pay for.

An on-device ADHD brain tracker for Android: 30-second voice entries in, sourced behavioral patterns out. No grading, no gamification.

Gemma 4 E4B takes native audio (no SpeechRecognizer) and streams transcription plus follow-up, then three background lens passes feed a deterministic Kotlin convergence resolver. EmbeddingGemma drives tone-word drift; a sealed NetworkGate and a verifyNoTelemetry Gradle task make privacy a build-time guarantee, not a promise.

An AI-assembled portfolio site built to test orchestration and authorship boundaries under human supervision.

A Next.js app indexing every project, principle, and post through Algolia, with an AI chat layer for conversational retrieval instead of static browsing. Every merge clears CI-enforced coverage floors and Lighthouse budgets — 100% desktop, 92%+ mobile performance — before it ships.

A published linter that hard-enforces AI attribution directly in commit workflows.

TypeScript and Python rules plug into commitlint and gitlint, enforcing explicit AI attribution on every commit that touches AI-assisted code. Mechanical and strict by design — a failing git hook, not a policy document nobody reads.

A lightweight pipeline that turns messy thoughts into structured System Notes index cards, stored durably in Supabase.

Exposed as an MCP server so any AI client can write directly: an Express API validates every payload against Zod schemas before Postgres ever sees it — enforced at capture time, not cleanup time. Freeform input in, structured System Notes cards out. No exceptions.

A reusable ESLint configuration built for AI-assisted codebases.

Echo collapses formatting and linting into a single shared config so the same boundaries apply across projects and agents without restating preferences. Built for AI consumers first, humans second.

A satirical incident-management app where Hotfix the dragon roasts your worst code screenshots — and thermally decommissions them.

React 19 and Tailwind front a PixiJS-animated dragon and Howler audio, while an Express API calls Gemini Flash to generate deadpan enterprise incident reports. Each report writes to a real-time Firestore manifest, powering a public leaderboard of P0 disasters.

An automated, AI-optimized mirror of my technical writing.

A Python job on GitHub Actions pulls my DEV posts, renders them through Jinja2 templates optimized for AI parsing, and deploys to Firebase Hosting. Not SEO or analytics — it exists so AI systems can correctly index and cite the work.

A free, sharable personality quiz that scores you across all five "vert" types — introvert, extrovert, ambivert, omnivert, and otrovert.

A deterministic, fully client-side TypeScript engine scores 35 questions across four axes into five independent fit percentages — the bars don't sum to 100. No backend, no tracking, no persistence. Built with vibes, AI, and a small pile of cited research.

A public signal-triage dashboard that routes volunteer attention to the DEV posts the feed would otherwise bury.

A Next.js app that pulls the Forem API, then runs each post through a GPT-5 nano-to-mini cascade scoring engagement and conversational signal — not recency — to flag high-need, low-interaction questions before they vanish beneath higher-performing content.

A personal AI-agent playground exploring memory as a first-class concern.

Python and LangChain orchestrate a local agent backed by TimescaleDB + pgvector for memory — Ollama runs locally, OpenAI handles embeddings. Built to watch memory behave, drift, and misbehave, without production pressure shaping the decisions.

Narrative travel discovery turned into an experiment in AI orchestration and parallel workflows.

Originally orchestrated in n8n, then rewritten as a Django API calling OpenAI for parallel AI workflows, deployed as a Python Worker on Cloudflare's edge — a rewrite that broke as much as it taught. Supabase and React back a discovery surface reused once cost constraints reshaped the project.

A curated, opinionated set of GitHub Copilot workflows, agents, and patterns that earned their place.

Not a neutral list. I build a lot of Copilot tooling locally and most of it dies there; only what survives daily use, frustration, and edge cases ships here. It reflects how I actually work with Copilot, not how the docs say I should.

A temporary AI coding agent shipped as a GitHub Action.

A GitHub Action calling the GitHub Models API (GPT-4o) to draft PRs under my own scoped PAT, with explicit human review before merge and prompt-injection detection built into the pipeline. A deliberate stopgap until Copilot's coding agent supported PAT-based CLI runs in Actions.

A falsified experiment in portable Copilot chat context across IDEs.

I tried to persist Copilot chat context across environments and found the API simply couldn't support it. Archived immediately — a fast, deliberate kill to avoid sinking effort into a dead end.