ABOUT · THE SYSTEM

One founder. A fleet of agents. One system that builds, ships and checks its own work.

Ghost AI Systems is Daniel Castillo, a U.S. Marine Corps veteran in Orlando, and the agent system he built to do a studio’s worth of work. This page is how it works.

See the pipeline

I don’t run an agency. I run a system. I design the work, agents gather and build, a second wave of agents tries to break it, and nothing ships until I’ve read what they found.

HOW THE SYSTEM WORKS

Gather. Challenge. Judge. Ship.

  1. 01 / GATHER

    Gather

    Small, fast models fan out in parallel. Each gets a narrow brief and returns structured data against a schema, not an essay. My GitHub audit this week ran as 19 batches across 105 repositories and came back with 322 projects.

  2. 02 / CHALLENGE

    Challenge

    A second wave argues the other side. Every “delete this” gets a skeptic hunting for a live deployment or a client who depends on it. Every “open-source this” gets one combing git history for leaked keys. This week they overturned 35 of 100 calls.

  3. 03 / JUDGE

    Judge

    A frontier model reads only what survived and makes the call. It doesn’t redo the gathering. It decides, and it cites the evidence it decided on.

  4. 04 / SHIP

    Ship

    I approve, and a push to main deploys production. Type checks and the test suite run first. Payments and live phone lines don’t get touched without tests.

$ workflow run repo-portfolio-audit› 19 batches · 105 repos + local projects · read-only› run 1: 8 agents live at once → host out of memory (16 GB)› run 2: MAX_LIVE = 5 · semaphore wrapped around agent()› gather:video-ai-flagships       model=small  effort=high› gather:website-core-ai-modules  model=small  effort=high› verify-delete → looking for live deploys, client ownership› verify-oss    → scanning git history for secrets› done: 50 agents · 322 projects inventoried · 0 errors› skeptics overturned 35 of 100 delete / open-source calls› judge: reads only what survived
An excerpt from the repository audit I ran on November 4, 2026. The first run crashed the machine, so the second one learned to queue.

Agents forget.
The system doesn’t.

Every agent session starts empty, so the memory lives outside the model: plain-text notes, one fact each, with why it matters and how to apply it, linked to each other.

Client decisions, pricing canon, which key lives where, what broke last time and why. A relay file hands work from one AI coding tool to the next, so whichever agent picks it up starts where the last one stopped instead of starting over.

No favorite
model.

Models get replaced the week a better one ships. Providers sit behind one router, so switching is a config change, not a rewrite.

Small, cheap models do the volume work. Frontier models do the judging. The voice agent on this site falls back to a second provider if the first one drops, mid-call. Which vendor is on top this month is a benchmark result, not a loyalty.

Counted,
not estimated.

COUNTED NOVEMBER 4, 2026

105
GitHub repositories, one account and one org
128
pages on this site, all in one Next.js App Router codebase
731
automated tests that run before anything deploys
04
model providers behind one router, one of them running locally
199
memory notes the agents read before they work
05
agents allowed to run at once. The 16 GB Mac set that limit.

The rules the agents run under.

  1. Read-only until I say otherwise.

    Audits look. They don’t push, delete or change visibility.

  2. No number I can’t defend.

    The figures on this page were counted on November 4, 2026.

  3. AI says it’s AI.

    Voice agents never pretend to be human. Being an AI is the point.

  4. Secrets stay in env files.

    Not in a prompt, a log, a screenshot or a commit.

  5. Client work stays private.

    What’s public is mine to publish. Client code isn’t.

  6. Production ships on my say.

    An agent can open the change. Only I deploy it.

Read the code.

Parts of the system are public. Fork them, break them, tell me what you find.

Point the system at your business.