I spent 3 months debugging why an AI voice agent kept booking appointments in the wrong timezone (and what it taught me about shipping AI pr
I've been heads-down building AI automation systems for small businesses — most recently a full lead-follow-up and booking system combining an AI voice agent, SMS, and a CRM.
The AI part was the easy part. The hard part was everything around it:
A calendar silently inheriting the wrong timezone from the account settings
A booking API that returns a misleading error because of one missing field
Making sure the voice agent and the SMS system never step on each other
None of that is "AI engineering" in the way it gets talked about online, but it's the actual work of shipping something a business can rely on.
I'm building out a freelance/agency practice around this kind of work — AI agents wired into real tools, not demos. If you're building something similar or know someone who needs this, I'd love to talk.
Happy to share more of the technical writeup if useful.
Great example of why the hardest part is often everything around the AI.
We’ve seen better results by adding layers of agents, almost like a company structure with junior and senior roles.
At first, humans confirm the agent’s tasks and decisions. But once the same type of decision is repeatedly approved, that confirmation can gradually be delegated to another agent layer.
Instead of removing humans all at once, we’ve found it more practical to reduce human involvement step by step as the workflow proves itself.