Drop in a Guitar Pro or MusicXML file and get it back with a chord stamped above every bar, a fingering diagram for each one, and a strummed backing track that follows the changes. The chord-detection engine reads every note in a bar and works out the chord in the song's key, so it labels the bars the tab's author left blank. Indexes a whole tab library in the browser. A Learn section for new players has interactive widgets (CAGED, Circle of Fifths, ear training) that stay in sync: set the key in one and every widget follows.
JavaScript, Python / FastAPI, SQL, Oracle Cloud, Cloudflare.

Runs 24/7 under systemd on a home server (Ubuntu). Gmail pushes new mail through Google Cloud Pub/Sub, Claude Haiku classifies it, the service applies the label before I ever open the inbox. Picked the smallest model that handles the job to keep token costs down.

Our D&D group plays online, and the campaign notes write themselves. A multi-track Discord recording bot exports each player's audio as FLAC; Whisper transcribes each track locally into one time-ordered, speaker-labelled transcript; Claude writes an in-character session recap from it, with guardrails against invented facts and spoilers; and the site rebuilds from markdown and deploys to Cloudflare Pages. It runs unattended as a systemd service, and the audio never leaves my machine.
Python, Whisper, Claude Code, Cloudflare Pages, systemd.

A ~20-page Django dashboard using the freqtrade REST API that monitors several bots at once, with health checks, auto-restarts and a read-only Claude Agent SDK assistant that talks through the data on any page. The research pipeline behind it runs hyperparameter optimization with walk-forward and out-of-sample gates, plus a causality auditor that rejects any signal peeking at future prices.

Skills written in the open SKILL.md standard that Claude Code and OpenAI's Codex both build on. One of them gets a person up to speed on a product or console they don't know yet: it drives the real interface in their browser, spotlights each control, sets a small task, then waits while they try it and ask questions. Not a video, not a docs page: the actual product, with someone riding along. It ships with a 19-scenario eval suite graded by code: without the skill, the agent acted for the user or typed a secret in 18 of 57 runs; with it, 0 of 57.

A local MCP server that lets the Claude app send messages for me: WhatsApp from my own number, Discord DMs from my bot. It asks before every send by default and never guesses between two contacts with the same name. Small on purpose, and instrumented like something bigger: structured logs, OpenTelemetry traces and metrics into Jaeger and Prometheus, and a doctor script that connects the way a client does.
Python, MCP, OpenTelemetry. 142 automated tests.

Eight years in technical support. At 1Password I handled escalations for enterprise customers and IT admins, reproducing bugs across Windows, macOS, iOS, Android and browser extensions, and digging through Datadog and Kibana at the API layer whenever a ticket gave me a reason to. Before that, nearly seven years at Applied Systems supporting insurance brokers on comparative quoting software, where I wrote a Python EDI service monitor that eliminated our biggest single call driver.