I am a second generation tattooer in Greenville, South Carolina. I own a tattoo shop, and I build the AI systems that run it: booking, messaging, agent orchestration, and an engine that trains those agents on real work and refuses to promote one until it passes a frozen trial.
Apex is the attunement layer in my Ark workspace. Live conversation telemetry from the shop is admitted through a reviewed capture source, turned into versioned candidate profiles, and tested in a sandbox against frozen scenarios drawn from real edge cases. An independent evaluator scores every scenario, deterministic gates check the rules a model cannot talk its way around, and a candidate is promoted only after a Crucible verifies the evaluation artifact and I approve it.
The numbers below are rows in the database, not estimates.
The blocked candidate is the part I am proudest of. At rubric 2.5.0 it scored 0.5, 0.5 and 1.0, failed the question budget and handoff rules, and never reached production. The gate did its job.
Stack: TypeScript, Hono, Drizzle, Postgres, Vertex AI, LiteLLM
The shop's public site and artist portal: booking, an event ledger for every message in and out, Google Calendar scheduling, and a card table where nothing that costs money or leaves the building moves without a ruling. React and FastAPI on Vercel, Railway and Cloudflare. Roughly 500 to 600 clients a year run through it.
Live
My AI operating workspace: agent routing, model gateway, capability registry, secret resolution, and an evidence registry that records what is actually built instead of what the docs claim. Everything above runs on it.
In service
I am building a public read only MCP server so you can connect Claude to this portfolio and interrogate the architecture yourself: pull the evidence behind every number on this page, read the case studies, and run sandboxed simulations of the real pipelines against snapshot data.
mcp.justinwalker.dev // status: building