Send me your worst agent conversation. I’ll tell you why it broke.

I spent years at Apple building conversational AI agents for production, the kind serving millions, where “the demo works” and “it works in front of real users” are very different problems. The reason an agent fails in production is almost never where you’d guess from the code.

So send me one real conversation — the one where your agent confidently said the wrong thing, looped, or “handled” a request it absolutely did not handle. I’ll read it turn by turn and send back the failure point you couldn’t see and the first thing I’d change. Free.

Customer-support, ops, coding, research agents — all fair game. Redact anything sensitive first.

How it works

1. You send a real (redacted) conversation by email — your worst one (or a batch, for the full walkthrough).
2. I read it turn by turn and find the actual failure point — the place it went wrong, not just the symptom you saw.
3. You get back — for the free teardown, the failure point and the first fix, in writing; for the $49, a recorded walkthrough over your traces: why each broke, the prioritized fixes, and where your post-launch setup is flying blind. Usually within 3 working days.

As part of the teardown I’ll also flag where your agent looked like it resolved the issue but actually deflected — restated the question, punted to “contact support,” or gave a confident non-answer. It’s the failure mode that hides inside a green dashboard.

The free teardown comes back in writing; the $49 is a recorded walkthrough — a video link, not a doc. It’s expert work, not software, so there’s no login or dashboard. Both land within 3 working days.

Your data & privacy

• Please redact PII first (names, emails, order IDs, payment info).
• I delete your transcript within 7 days of delivering your teardown.
• I never reuse it or share it. It goes to one place, nowhere else.
• Payments are processed securely by Creem (the merchant of record). I never see or store your card details.

Refund policy

If the walkthrough doesn’t give you something concrete to act on, email me within 14 days of delivery and I’ll refund you in full. It’s custom manual work — I’d rather you be genuinely satisfied than keep a sale.

I’m Yaqin Hei, ex-Apple AI engineer. I write about the gap between agent demos and production — including the resolution-vs-deflection problem behind this teardown: Containment rate vs resolution rate.