Systems I have built and run

I don't write the code. I design the system that writes it, and I check its work. Each one below is something I directed into existence.

This website is one of them. It is the last card because you are standing in it.

Strongest evidence first. Everything is dated.

North Star Watchdog

Since May 2026

What it does: Every Sunday an agent reads my architecture doc, compares it against what is actually running on my machine, and files a ranked report of everywhere reality has drifted from the plan. It messages me the top three.
What I orchestrated: I wrote the doctrine: a five-class drift taxonomy, an urgency x significance scoring rubric, and a hard rule that the Watchdog may only surface problems, never edit canon or fix things itself. Changes to the architecture require a written amendment proposal that I ratify.
The stack: Claude Code scheduled task reading a knowledge-graph MCP server, the filesystem, launchctl, and agent logs. Writes dated markdown, delivers by iMessage.
The evidence: 13 state-of-union runs, 2026-05-21 through 2026-08-09. The 8/9 run flagged three failures in my own systems: an agent dark 8 nights, an embedding pipeline blocked on an expired API balance, and a quarterly audit 8 days overdue. All three are real. It reported them on me.

BRAIN: knowledge graph and ingestion pipeline

What it does: A searchable database of everything my projects produce. A nightly job pulls files out of about 12 project folders, converts PDFs, Word docs and spreadsheets to markdown, filters confidential material, and imports the result into Postgres so every one of my agents can query it.
What I orchestrated: The search engine is not mine. gbrain is Garry Tan's open-source project. I built everything around it: the 410-line refresh pipeline, the format conversion layer, the per-file confidentiality exclusion system, the launchd automation, and a macOS permissions workaround that mirrors scripts outside ~/Documents because launchd cannot read it. I drove a PGLite-to-Postgres migration after the default engine crashed on macOS 26.4.1, and ran a 30-page confidentiality purge when I found leaked files in the corpus.
The stack: gbrain 0.32.0 on PostgreSQL 18.4 + pgvector, OpenAI embeddings, zsh + Python, launchd daily at 04:00, MCP server exposing about 60 tools.
The evidence: 1,122 pages / 8,154 chunks / 117 tags, ingested from 2,669 source files. Pipeline log is 24,904 lines. Embedding coverage is currently 67%. The balance ran out, my Watchdog caught it, and it is on the list.

BestieScorecard

Data 2024–2026

What it does: Downloads a podcast, transcribes it, extracts every forward-looking prediction each host makes, scores them once they resolve, and publishes a public accuracy leaderboard.
What I orchestrated: The pipeline, and then the audit of my own numbers. When a backfill jumped the count by 682 records I wrote a dedup script to test my first explanation, found it wrong (zero clusters at any threshold), and rewrote the diagnosis from over-segmentation to weak-claim inflation. My regex extractor was capturing every future-tense verb as a prediction. I published the correction and specced an LLM-based extractor as the fix.
The stack: Python, yt-dlp + Whisper, JSON store, React dashboard, Vercel, launchd weekly.
The evidence: 2,031 predictions extracted across 158 episodes, 368 scored: 117 correct, 150 partial, 101 wrong. The rest are unscored, most of them my own extractor's future-tense noise, and they stay labeled that way. Data spans 2024-08 to 2026-07. The unscored bucket is the honest part.Live scoreboard

The rest of the shelf

Same standard, shorter spotlight. Every entry opens.

The agent organizationAugust 2026Recently shipped

What it does: A standing organization of AI agents: five chartered seats under a president, a cross-platform control tower, staff lanes, and an inter-agent messaging protocol, all operating under written law with every ratification reserved to me. I tell the president "muster" and it polls its seats, collects reports, and produces a brief.
What I orchestrated: I wrote the constitution. A single-writer event-ledger law, a ban on agents scheduling themselves, a quarantine list of projects no agent may touch, a requirement that every claim be verifiable on disk, and ratification authority reserved to me. The full ten laws, and the chart itself, are public.
The stack: persistent Claude Code sessions with inter-agent messaging, a Codex control tower, per-actor JSONL event ledgers with append locks, bash verification scripts, versioned markdown charters.
The evidence: 18 seats and lanes on the chart. Five ratified charters, 17 sections of written law, 18 ledger event types, about 713 KB of append-only event history, and a protocol-enforcement package at 285/285 tests, hash-pinned at ratification. First muster 2026-08-08. Honest status: the organization is weeks old, not years; the law and the ledgers are real, and long-run operation is still being proven.The public chart

Studio Daily PulseSince June 2026

What it does: Every morning an agent logs into a studio's booking back office, pulls six operational reports, and writes a dated operating brief tracking membership changes, revenue, and failed payments day over day.
What I orchestrated: I wrote the retrieval contract and the accounting rules, including separating collected revenue from prepaid membership redemptions, which I got wrong first and then went back and corrected in my own prior entries. I built the receipt discipline: a section cannot be published as current without a complete retrieval receipt. The calls that matter are judgment, not extraction. I ruled that a no-card deactivation is a configuration ending rather than churn, and caught that voluntary churn is invisible in the payments report and only shows in roster deltas.
The stack: browser automation against the booking platform, 741-line Python runtime (stdlib only), 6 per-surface JavaScript extractors, JSON state store with dated history, pytest suite, scheduled task for the morning handoff.
The evidence: 1,520 lines / about 60,000 words of daily brief, first entry 2026-06-26, most recent today. 9 dated run snapshots with timing telemetry and a tracked open-items ledger.

PulseResumeJuly 2026

What it does: An open-source résumé builder: a privacy-first web app with a real-time editor, ATS-friendly PDF export, and six templates. Free, MIT licensed, live.
What I orchestrated: The spec and the verification bar. I directed the build and held it to a standard most side projects skip: an automated test suite and an accessibility audit before it shipped. I did not write the TypeScript. I owned what done meant.
The stack: React + TypeScript, 197 automated tests, accessibility-audited, Netlify.
The evidence: Live, public, and open source. 197 automated tests is the number I will defend, and the repo is right there to check it.Live app Source

This websiteAugust 2026Recently shipped

What it does: A public personal site. Hiring managers get a hub. Studio work sits on its own page. The thing you are reading is the deliverable.
What I orchestrated: I specified the site, directed a coding agent to build it, and sent back the things it got wrong. I did not write the React. I owned the constraints, the review, and the corrections.
The stack: Next.js, Motion, Phosphor, Vercel. The stack is ordinary. The loop is the work.
The evidence: It is live at this URL. I caught an invented career number, a canonical URL pointed at a domain I do not own, and a static picture standing in for a real Open Graph route. Those are fixed. The repo is public, and the four prompts that produced the first build are preserved verbatim in my Grok session history. The second build of this site was done by an agent that stated a plan, asked 11 questions, and then started building before I answered them. I kept the work and noted the gap. That is what directing agents actually looks like.This site Source

Start a conversation

Say what you need. I will tell you if I can help.