Shipwright
- What it is
- An agent that takes an approved GitHub issue and returns a tested, reviewable pull request.
- Read first
- The agent runner's test suite, which exercises the plan, patch, and review loop end to end.
Three ledgers, checked in your browser: repositories you can read, demos you can open, and the private systems behind the case studies. Deeper backend checks live in the demo operations page.
Every entry names the file worth reading first. Entries tagged cold-clone reviewed were verified by an independent engineer working from a fresh clone before the repository went public. Claude plugin packs and demo-only repos stay on the GitHub profile and the live-demo ledger below.
Grades a codebase's technical debt and prices the paydown.
Turns AI coding-session transcripts into durable, searchable lessons.
Scores and routes inbound leads; a 24-case model eval suite gates every accuracy change.
A terminal dashboard for running a fleet of AI coding agents in tmux.
Fans one prompt out to a fleet of AI coding agents in tmux panes, then collects their work.
One CLI that routes a prompt to Claude, Codex, Cursor, or Gemini, with JSON output and real exit codes.
PostgreSQL fuzzy matching for public filings, with hermetic SQL assertions and a bounded live demo on Texas open data.
An autonomous coding agent that works a Linear backlog end to end and opens the pull requests.
The sanitized public kit behind Throughline: one four-method connector contract, the sync engine, sample CRM data, and tests.
One Zod-validated operation registers as an HTTP route, a CLI command, and an opt-in MCP tool, with OpenAPI emitted from the same registry.
Reads, injects, and exports 1Password secrets from the command line so they never have to sit in a committed .env.
Scores Claude prompts against a stored eval suite, with string-match or rubric grading you can re-run.
A Claude Code stop hook that sends the agent back to work when the last turn still lists next steps, and stops when it is asking you a question.
Forty-nine MCP tools over Meta's Marketing API, with a smoke test that starts the server and lists every tool without credentials.
Compiles a JSON show definition into a vMix preset bundle a studio machine can open.
Suggests the next Claude Code skill from a local library using fuzzy search, TF-IDF, and optional embeddings.
One command definition writes Raycast, PopClip, Dropzone, and Shortcuts artifacts, and will not emit a surface that cannot represent that command's input.
Drives Behringer X Air and Midas MR mixers from the terminal, and a parity test fails the build if the docs drift from the CLI.
Claude Code plugin packs for building, testing, and deploying, where CI runs the plugin validator over every pack on every push.
Resolves llm model keys from 1Password at prompt time. refs.json stores the op:// reference; a failed read aborts the call.
The rest of the public estate lives on the GitHub profile.
The production systems stay private; their case studies name what each claim rests on. The repositories above show the same engineering in the open.
Case studies with evidence basis and self-reported labels
Checks 6 live demo servers, plus the separately owned RevOps Factory demo
If a row here is the reason you want to talk, email me.
Senior IC · Dallas–Fort Worth · remote preferred or DFW hybrid