I run a watch & jewelry shop in Cleburne, Texas, and I build software in the margins around it. The through-line in everything I build is trust infrastructure — I've been scammed and messed over enough times that I build systems where a customer can't easily be messed over.

• dfwmetro.net — a Dallas–Fort Worth directory of genuinely online-bookable service businesses, with a verification desk (license, registration, insurance checked before a business gets the verified badge). Its MCP server is live and listed in the official MCP Registry as io.github.cexplayer01/dfwmetro-directory. The open-source stdio bridge is here: dfwmetro-mcp (MIT).
• Primed Harness (Bounded Agent Harness) — governance for long-running AI agent work: budgets set up front, explicit authorization for irreversible steps, cost known before a run starts, continuation/resume when a run hits a wall, and an audit trail of what actually happened.
• USB — white-label websites for local businesses ($20 down + $5/mo), built so each site doubles as a pre-verified directory listing with booking. Status page runs on GitHub Actions: usb-status.
• entropy-dice — random numbers for AI that never come from the model: layered, publicly verifiable noise (NIST beacon, drand, Bitcoin block hashes, live weather, timing jitter) conditioned through SHA-256.
• texas-pmtiles — a self-hosted basemap of Texas as a single 562 MB PMTiles file, so a map product never dies because a free tile service changed its terms.

Hard-won, documented lessons from shipping this stack solo:

• MCP servers: registry listing, stdio ↔ Streamable HTTP bridging, keeping tool definitions live-synced instead of copy-pasted.
• AI agent harnesses: per-agent governance adapters, budget ceilings with mid-run polls, verification at the layer where the behavior actually lives (not where the report says it lives).
• Supabase/Netlify traps: shared-project auth emails falling back to the wrong product's domain, deploy manifests silently dropping live files, and why I diff every production deploy before/after.
• Programmatic-first workflows: scripts, CLIs and free agent routes doing the volume; expensive model tokens reserved for judgment calls.

I don't fork-and-forget. When I fork, it's because I'm bringing a fix or a feature I needed myself and proved on my own stack first — with tests, and a short note on where the problem bit me in production.

• Directory / MCP: https://dfwmetro.net
• Harness: https://boundedagentharness.com
• Shop: Van Daele Watch & Jewelry, Cleburne TX