Writing
Posts on building AI that's reliable enough to put in front of customers — alignment, auditability, and the engineering you don't see in the chat window.
"Just give us your number, we'll handle the Meta side" is the most consequential decision in the whole engagement, made in about four seconds. What a WhatsApp Business Account actually owns, where the patient conversations really live, what an exit costs — and the five questions to ask on the onboarding call.
A reschedule is a write to a live operation. Why a reschedule and a cancellation are different risks, why the scoped handover link is the design rather than a cop-out, one authority on what may move, and why stale intent has to expire on its own.
When one thing a customer books is several things the operation runs, the booking has to be the unit that stays consistent, partial success has to be treated as the normal case, and the seams underneath have to be genuinely closed rather than hidden in the wording.
Most of what looked like logic turned out to be a description of one client's catalogue, which is what separates the engine that gets reused from the configuration that never should be.
Scoped actions instead of open API access, why "it's just logging" doesn't stay true, documenting accepted risk instead of leaving it invisible, and why the pass belongs before you widen the surface, not after.
An assistant that is wrong is not obviously wrong. The failure classes that show up months after go-live, why a demo cannot surface them, and how to find out what yours has actually been telling people.
Two ledgers, one truth. Picking an owner and meaning it, the discriminator problem nobody warns you about, building for the second webhook, and writing the minimum a patient can be identified by.
The volume-vs-judgment split that actually decides where AI should act alone, why the boundary has to be an explicit rule, and how to make a handoff feel like an upgrade instead of a dead end.
Why a transcript isn't an audit trail, what a log record needs to hold, and how logging every answer and action is what makes accountability real rather than a slogan.
The jump from an assistant that answers to one that books, reschedules, and cancels on live systems: source-of-truth discipline, scoped authority, and safe, logged actions.
Using agentic methods to keep AI answers compliant — extending the "built to be trusted" line straight from the medical-group case study.
More posts to follow, including a write-up of the outreach build once it's further along.