The constraint

I had a procedure and was handed a photocopied sheet and a phone number. The instructions mattered — get the prep wrong and the procedure is cancelled and rebooked weeks out — and they arrived in the least reliable format there is. Worse, no two sheets agree: the timings, the products and the restrictions differ by clinic and by procedure, so there is no single correct plan to ship.

What I built

A prep companion for iOS and Android that turns that sheet into a timed plan.

• You photograph the sheet. The images go to the backend, an AI parse returns a structured instruction set, and the app builds a schedule counted backwards from an appointment that is different for every person. Asking someone to type their prep in would have reproduced the problem I was solving.
• Check-ins during the prep itself, including a clarity scale, so the app can tell someone they are on track — which is the actual anxiety at two in the morning.
• An affiliate and attribution engine for clinic and practitioner referrals: a link fingerprints the click, records it, and redirects to the right store.

Decisions worth defending

• Deterministic where the platform allows it, probabilistic where it does not. Android carries the referral code through the Play Install Referrer, so attribution on first launch is exact. iOS has no such channel, so it falls back to fingerprint matching or the user entering a code. Same system, two honesties — and the reporting has to admit which one produced a given match.
• The attribution engine is the one I built for WunderGuide. Same problem, same shape, second deployment. For one person maintaining several products, the reusable piece is not the UI — it is the infrastructure nobody sees.
• Operable by one person, deliberately. Every choice — store ops, billing, support load — was tested against whether I could still run it in six months while working on something else. That rules out a lot of otherwise good ideas.
• No marketing spend. Organic search and app-store discovery were the only acquisition channels, which keeps you honest about whether the listing answers what someone is actually searching for at eleven at night.

Operating it

The unglamorous parts are the product. Uploading several megabytes of photos and waiting on a parse takes longer than a default HTTP timeout allows, so the client has its own; a scan endpoint that accepts images from anyone needs replay protection; and an attribution ledger is only useful if you can tell a real referral from a re-clicked link. Shipped and operated solo, alongside full-time work.

Where it stands

Live on the App Store and Google Play, still operated solo.