Two Systems, One Caller: The Architecture Behind the AI Receptionist
We published the full engineering case study for the platform. Not a feature list — the mechanisms, the code, and the failure modes we designed out.
The äußern receptionist answers a real telephone, holds a real conversation, and books real work into a real calendar. The part that is genuinely difficult to build is not the language model. It is everything around it.
The receptionist is programmed in one platform. The calendar it books into is owned by another. The two share no database and no codebase — they speak only over an authenticated HTTP boundary, on purpose. Almost every hard problem in the system follows from that one decision, and so does almost every guarantee it can make.
Most writing about AI receptionists describes outcomes. This document describes mechanisms. It quotes the running system: the signed write, the revision counter that arbitrates barge-in, the lock that settles two callers reaching for the same hour, the supervisor configuration that keeps a headless browser from ever competing with a live call. Where something can fail, it says what happens instead.
It runs eighteen pages with five diagrams. Read it in the browser, or take the print edition with you.
Two hosts on one private segment, three authenticated channels, and a machine credential that renews itself before anything waits on it.
Seven stages that turn a business’s own website into a receptionist that already knows the business on day one.
Twenty ordered sections, and the invariant that keeps one compiler serving every trade without a single profession branch.
What happens when a caller talks over the reply while a write is already crossing the boundary to a live calendar.
Two callers accept the same slot in the same instant. One wins, one loses honestly, and the database decides — never the timing.
Two supervised worker groups, the CPU priority that keeps a headless browser off a live call, and the one integer that scales it.
Two Systems, One Caller
An autonomous receptionist holds a live telephone conversation, negotiates an appointment, and commits it to a calendar owned by a separate system across an authenticated HTTP boundary.
Code excerpts in the document are taken from the running system. Comments are verbatim.