They Declined the Meeting. I Built the App Anyway — Local-First, No Cloud, No Stored Health Data

Seventeen days after a hospital system declined to evaluate a patient-prep architecture, the working reference build shipped. Here is the engineering case, and the white paper that specs it.

On September 27 I pitched a hospital system an imaging patient-preparedness tool. In October the sourcing office declined — without evaluation of the architecture. The record is private; the pattern is public. On October 22, seventeen days after the decline, the thing existed anyway: Radiology Companion™ v5, a working local-first reference build. This post is the engineering case for why it is built the way it is built, and why I wrote a whole white paper to spec it before asking anyone to trust it.

The bug class nobody audits

Here is the failure, in engineer's terms. A hospital sends prep instructions through a portal. The portal logs sent. The patient experiences nothing arrived. Fifty-eight days after the message went out, an appointment died because of it — mine, and I have the receipts, because I document the way Ida B. Wells kept records: date everything, name everything, record the reasoning.

The system's ledger says sent. The patient's ledger says lost. Only one of them gets written down. A sent-flag verifies delivery to a server; arrival at a person is the other end of the wire, and nothing in the chain checks it. It is a behavioral control that fails silently — that is what behavioral walls do. Missed US medical appointments get quoted at $150B a year; that figure is trade press and it travels flagged in everything I publish (the peer-reviewed clinic-scale literature backs the pattern: roughly a quarter of patients run nonadherent to care that includes keeping the appointment, and the prep-failure chain in endoscopy is measured in lesions and repeat procedures — receipts in the paper).

The build

Seventeen days, one builder, AI collaborators, no funding. The constraint set did the design work:

• No cloud. No backend to breach, no accounts to provision, no data to sell.
• No stored health data. Nothing about the patient leaves the device. Not analytics, not logs, not crash reports.
• The protocol explained in order. Every preparation step verified, every step read aloud.
• Twenty languages, because the register problem is also a language problem.
• A hospital's own name and logo, fetched from a pasted link — the only configuration the institution supplies.
• Four sourced institutional demo libraries — the prep pages Mayo, Johns Hopkins, and Cleveland Clinic already publish, cited, never generated.

That last point matters. The app invents no prep. The canon is cited; the app is the delivery layer that gets the sheet out of the app: paper, calendar, read-aloud. Exits before entrances. The department's sent-flag measures nothing that matters; the companion's whole job is to make obedience to any single channel stop being the test.

Why "local-first" is the requirement, not a preference

Health data is the most sensitive class there is, and the pitch that got declined was a 90-day fixed-fee pilot against the institution's own reschedule baseline. A system that holds nothing cannot leak anything. A demo that runs on a laptop needs no procurement cycle. The institution's only integration is pasting its own URL. The architecture assumes the deployment is the customer's — which is exactly why the source is held: demo free, deploy paid; source private; commercial license only; review under NDA. The party that owns the record owns the terms of contact.

The white paper is the spec

Before asking any department to trust the app, I wrote the paper that justifies the need: The Class of the Prep Sheet — What Bourdieu Teaches Imaging Departments About Failed Preparation.

It is a record-paper, and it verifies its own claims in the open: every clinical citation resolved to full form (authors, journal, volume, pages, DOI) and checked findings-as-used at page level; the one trade-press figure carried with its flag showing; no physician, staff, or facility named; the first patient is the author. The frame is Bourdieu's — the prep instructions as an exam testing three kinds of cultural capital at once — and the pedagogy chapter names the mechanism with Freire's: instructions deposited in a portal, adherence graded at withdrawal, banking education wearing a scheduling badge. The fix has been in medicine's own literature for twenty years. It is called teach-back. The sent-flag closes no loop; teach-back closes the only one that matters.

The return date

The paper files the design requirement; the capability build is my move and my timeline. The reference build is demoable on request. If your department has a reschedule baseline and a prep-failure number, the pilot design is already shaped: 90 days, fixed fee, measured against your own figures — the party whose performance is measured should not be the party holding the measuring instrument.

I build AI tools that solve everyday problems. What problem can AI and I solve for you today?