AI engineer · IAM · Author
I build AI tools — a kernel that helps agents check their own work, a writing app that remembers your story, a custom MCP server. Identity & access on the day job. Writing WESTFALL in the margins.
agent-loop commands · BrainOS · arena-tested
custom MCP server · ReVerse drives UEFN
recommended picks · Righteous Recon · 44 sectors
verifiable tokens · Northbridge · live
01Headliner
the kernel that helps an AI check its own work
A small kernel that keeps AI agents honest. It picks where to look for answers, tracks what was claimed against the source it came from, names what counts as proof up front, and only closes a task when the work actually checks out. Eight commands cover the whole loop. Running stress and arena tests this week.
A small IAM sandbox where everything works the way it would in production. Real tokens you can verify, audit logs you can't quietly tamper with, revocation that lands in about a second, two-person approval for the risky stuff. There's a /redteam page that's honest about what's real and what's faked — so reviewers don't have to guess.
02The shelf · shipped & in flight
03What I do
AI engineering — Small tools that help agents check their own work before they ship it. Custom MCP servers when an LLM needs to drive a real piece of software. Claude API calls inside the apps I ship — usually the cheap model, the expensive one only when the work calls for it.
Software shipping — Astro and React in production. Supabase auth, Stripe billing, Cloudflare Workers deploys, content validated by Zod at build time. I keep the runtime simple — most pages are static, real state lives where it has to.
Identity & access — Entra ID lifecycle, access reviews, on-demand admin elevation, ticket triage, and onboarding for production teams. Security shows up in everything I ship — not a layer on top, something I start with.
04The ledger
github.com/zgib89 →A small kernel that keeps AI agents honest — picks where to look, tracks what was claimed, names what counts as proof, and only closes a task when the work checks out. Running stress tests and arena evals.
A writing app for novelists. v1 is live; the v2 rebuild — type naturally, every character and timeline event tracked behind the scenes, and a single fact change shows exactly which pages need to update before you commit — is in flight now.
An IAM sandbox where everything works the way it would in production. Real tokens you can't forge, audit logs you can't quietly tamper with, revocation in about a second, two-person approval for risky operations, and a page honest about what's real vs theatrical.
Make a real portfolio site in 90 seconds. Pick a template, edit, hit export — out comes a complete project you can host anywhere. No AI runtime, no account, no lock-in.
A book-first conspiracy thriller told through evidence: the novel, a case-file site with public dossiers, a soundtrack lab, and direction for a top-down game adaptation. Conscience costs everything.
Honest gear reviews across 44 sport sectors. Files-in-folders content, schema-validated at build, static search. Nothing is recommended that wouldn't pass the 80/100 threshold.
A custom MCP server so Claude Code can drive UEFN like a CLI — placing actors, editing scenes, running operations the editor normally makes click-by-click. The LLM is the controller; the side effects are typed tools.
05On the desk
06Say hi
Building something where AI, identity, or story collide? I read everything that lands.