I built your intake specialist.
Talk to her right now.

The Network Financial voice agent above. It qualifies the caller, calculates real appointment times, and hands off to the CRM — end to end. I designed the conversation flow and the architecture. Claude Code wrote the handler to my spec.

Similar routing logic runs in client accounts — the round-robin lead assignment linked below is mine end to end.

8 GHL accounts, 2 agency-level — real estate, coaching, auction, e-commerce, HVAC, and solar. High-ticket lead routing with round-robin assignment. SMS AI agent. The agent above: built today, after I read your email reply.

GHL 8 · Vapi 8 · n8n 8 · ElevenLabs 8

Built a full GHL system for one client in 2 days after studying their sales process. The voice agent above took one.

Week 2 is watching, not building. Weeks 3–6, fix what real usage exposed — production always teaches you things you couldn't predict. Weeks 7–12, shift from fixing to improving.

Automations decay quietly. Nothing goes live and gets forgotten.

I alert on silence, not errors. A workflow that fails loudly is easy to catch. One that quietly stops triggering is what costs money — so I watch volume drops, not just failures.

Flag the wrong bookings first so the team can call those leads back. That comes before the fix.

Then isolate before touching anything. Timezone bugs live at the seam between systems, not inside one. Fix it in a single place, verify at the edges where offsets break, and make sure there's one source of truth instead of four.

Weekly: what's moving and what's stuck. Monthly: what's decaying — unused workflows, expired tokens, duplicate contacts, compliance status.

Half of GHL ownership is deleting what nobody uses anymore.

Dashboard as source of truth, plus a Monday summary: what moved, what broke, what I'm changing, one recommendation.

I'd lead with cost per booked appointment. That's the number that connects everything else to revenue.

Project work means you build it, hand it off, and never see what happens after week 3.

The first build is a guess. Months 2 through 6 are where it becomes reliable — and that's the part I want to be there for.

The strongest control is scope, not rules. The agent qualifies and books — it doesn't advise. It can't say the wrong thing about something it never discusses.

On top of that: hard-coded disclosures, a deny list, every call transcribed and reviewed, and no prompt change goes live without a test call.

The greeting stuttered on cold starts. My first fix made it worse because I changed a setting on a hunch instead of measuring.

Small problem, but the lesson was the pattern: I acted before I isolated. That's the mistake I don't want to repeat on something bigger.

Run the whole system and find where it actually breaks, not where I assume. Most problems get misdiagnosed because people fix before they isolate.

Once I know which layer is broken, I use Claude Code to fix it, then test until the output matches what it should be. And I don't go quiet. You'd hear the same day what I've ruled out.