Skip to content
View jaysueno's full-sized avatar

Block or report jaysueno

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
jaysueno/README.md

Jay Sueno

I learned to code at forty. Now it runs the company.

I'm the founder and CEO of Vultun, a wellness group of ten business units on Siargao, a surf island in the southern Philippines. I'm not a software engineer. I'm an operator with a data science certificate and a lot of agents. I design the software that runs the company, review it, and merge it. AI agents write nearly all of it. I'm forty-seven, seven years sober, and I started surfing at thirty-three. Still paddling out.

Almost all of that code is private, because it is full of real people. I'm pulling out the parts I can share, one small piece at a time.

Jay in an orange shirt crouched low in the pocket of a steep green wave, with white water behind him and a low island on the horizon. Bomi, Siargao. November 16, 2024. Photo by Tom Soupart.

Where it started

My first code was BASIC on an Apple IIGS, a machine that booted straight into a prompt and ran whatever you typed. Then Sierra's King's Quest showed me a computer could hold a whole world, and in 1997 I was one of the first players in Ultima Online, where a hundred thousand of us found out what a shared world does to people. Then I started businesses. The thread is the same the whole way: technology meets the real world and real people, and something happens that neither does alone. That is the magic sauce. It is why a wellness company on a surf island runs on software I build.

2026, in commits

Counted November 7, 2026, across sixteen private repos.

Heatmap of 2026 with one mark per day from January 1 to November 7: 4,305 commits across 216 active days, a longest streak of 39 days and a busiest day of 91 commits.

Four tiles: 4,305 commits, 3,168 merged pull requests, 84% with a Claude co-author line, and 216 active days of 301, counted November 7, 2026.

4,305 commits on main branches, January 11 to November 7
3,168 merged pull requests
84% of those commits carry a Claude co-author line. I read, I decide, I merge.
216 active days out of 301. Longest streak, 39 days. Busiest day, 91 commits.

What I'm sharing

Real operator software that runs a real business, one small piece at a time. Each one opens with the dated mistake that caused it.

Repo What it is Status
claude-operator-kit The hooks, rules, and skills I run every session with. Each rule carries the date of the mistake that caused it. public
agent-cage How I keep a personal AI agent in a box: read-only filesystem, default-deny network, and the fire drills that prove it holds. cleaning up
ollama-context-guard One model, one context window. The outage, the measured cause, and the guard that keeps it from coming back. public
llm-eval-gates Evals that block a model or prompt change when it regresses. Hard-zero gates, harnesses that prove themselves first. cleaning up
markdown-brain A second brain in plain Markdown, filed and recalled by a local model. The machinery, never my notes. cleaning up
sealed-tests The grader writes the tests first and seals them. The builder cannot touch them. cleaning up
autoresearch-mac-mini I let an agent tune my Mac Mini overnight. The best change was cutting my system prompt to one line. public
creator-ledger Track your own posts and audience across platforms with a few scripts and three CSV files. cleaning up
email-bounce-breaker Two small libraries that pause a send when hard bounces climb. Born from a suppressor that ran dead for four months. public
jaysueno.com The code behind this site: plain files an agent can edit, a Markdown twin per page, llms.txt, a lead form behind row-level security. public

How I build

Five steps from left to right: Jay designs and reviews, agents write, a deterministic gate checks, a human approves money, messages and health data, and the live object is verified.

  • Functions decide and fetch. The model phrases.
  • Every rule carries the date of the mistake that caused it.
  • A human approves anything that touches money, customers, or health data.
  • Verify against the real thing, never against the model's claim.

What I use

Tool What it does
Claude Code Where nearly all of it gets written
Mac Mini with a local Qwen model The repetitive work, on my own hardware
Supabase The database
Resend Email
DigitalOcean VPS Runs the crons and the agent
Astro and Next.js The websites and the internal app

Before the island

University of Washington. Finance at Boeing. Commercial property and a multi-family development in Olympia. Co-founded MamaP in Seattle. Back to school at forty for a data science certificate from UC San Diego, then data analyst at Amazon. Moved to Siargao in 2022.

Elsewhere

jaysueno.com · The Long Paddle · X · Instagram · LinkedIn · Resume · resume.json · llms.txt

Pinned Loading

  1. agent-cage agent-cage Public

    A Docker cage for a personal AI agent on a Linux VPS: read-only root, every capability dropped, default-deny egress, a guard that keeps the rules above Tailscale's forward chain, and five fire dril…

    Shell

  2. claude-operator-kit claude-operator-kit Public

    Two Claude Code hooks, six rules and four skills for running several AI coding sessions on one machine and one working tree. Each rule opens with the date of the mistake behind it.

    Python

  3. dawn-patrol dawn-patrol Public

    A small Python pipeline that pulls the Open-Meteo marine and wind forecast for three surf breaks, stores every pull in SQLite, and publishes a static page with a swell chart and a three-line brief.…

    Python

  4. llm-eval-gates llm-eval-gates Public

    Evals that block a model or prompt change when it regresses: hard-zero gates for invented values, baselines that only ratchet up, harnesses that reproduce their own pinned numbers before they score…

    JavaScript

  5. review-desk review-desk Public

    A model drafts the reply to a review, a person approves or edits it, and every edit becomes data: few-shot examples for the next draft and, when it repeats, a rule every draft must pass. An eval ov…

    JavaScript

  6. studio-in-a-box studio-in-a-box Public

    Studio software for an invented yoga studio: bookings, package credits kept as a ledger that must balance, an end-of-day cash close, and staff roles enforced in the database with row-level security…

    PLpgSQL