DeHoardMe.

Who it’s for

It’s for people clearing out their stuff who get stuck on what to keep. The testers I planned the beta around were people who complain about clutter, people who have just moved, and people doing an estate cleanout. I built it for them. I haven’t validated it with them.

The idea goes back to a couple of years in the trades, and what I saw on restoration jobs. It’s a fun side project, and it sits behind 50RH and Lantern.

What I built

You photograph something you own, and the app gives you a verdict, keep or toss, with a one-line quip in one of three voices: The Brutalist, The Therapist or The Nonna.

• Sleep on it. Save for Later puts a verdict on hold, and the app asks you again seven days later.
• History. Past scans with thumbnails, filtered by Keep, Toss or Pending. Any verdict can be shared as a card.
• Pricing. A new account gets 10 free scans. Pro buys one thing, unlimited Standard scans. History, Save for Later and all three voices are free to everyone. The release notes say plainly that Pro is thin, and I didn’t move features behind the paywall to make it look fuller.
• Safety. Before your first scan, you accept a one-time content agreement that includes confirming you’re 17 or older. Every photo is screened for content before the AI sees it, and flagged images are never stored.
• Expert scans, shipped dark. The server side is built, but the app can’t request one, and the server flag was off on 4 October 2026, the last time it was checked.

It’s a SwiftUI app on a FastAPI backend, with Supabase for sign-in, data and storage, and RevenueCat for subscriptions. I wrote it with AI, and the calls on the product, its pricing and what goes into each build are mine. Of the 64 commits on main, 59 have Claude as author or co-author, read 5 October 2026.

The first TestFlight build came about three weeks after the first commit, with commits on only four days in between, and the app kept changing through Build 15.

How fast

1. 25 MarFirst commit: a written plan for a photo-to-verdict decluttering app.
2. 15 AprFirst TestFlight build: the app installs on a phone and signs in with Apple.
3. 28 AprSave for Later: sleep on a verdict, and the app asks again in seven days.
4. 16 MayBuild 14: every photo is screened for content before the AI sees it.
5. 3 AugThe payment and free-scan paths get tests, and server errors are now reported.
6. 1 SepWith no sold-price data from eBay, Expert became a “worth listing?” call, and Pro+ came off sale.
7. 17 SepBuild 15 on TestFlight, with notes correcting what Build 14 promised.
8. 4 OctFrozen: maintenance only until I decide to ship it or retire it.

How it stays right

No invented dollar figures. Expert, the paid scan, was first built to return a resale value. When eBay refused sold-price data, I took the value out rather than let the AI make one up. Expert now gives a coarse call instead: worth listing, maybe, or not worth listing.

Measured before it’s sold. An Expert scan costs about 15 times as much as a Standard one, so it has to prove it’s worth that. An accuracy harness grades its “worth listing?” call against eBay sold prices collected by hand, and Expert can’t be sold until three gates pass: at least 40 shippable items scored, a serious-error rate under 10%, and a win over the cheaper model on the same prompt. A gate it can’t measure reads NOT MEASURED, never PASS. None has run yet, because the hand-priced set they need hasn’t been collected.

Corrections in public. When a promise stopped being true, the next build’s notes said so. Build 15’s notes withdraw three of Build 14’s “coming soon” claims, and the Build 14 notes stay as they were published.

Tests where the money is. The paid-access and free-scan paths had no tests until 3 August 2026, when one pull request added 25 and took the backend suite from 40 tests to 65. It was at 218 when the Expert rework merged on 1 September 2026, by the count that pull request quoted, not a re-run today. One test fails if the repo’s own docs deny a check that the payment webhook actually makes. Since 17 September 2026, the rule is that every pull request that changes code gets a review from a second AI model before it merges; a change to docs alone is skipped, by design.