Own productiOS and Android
Snixt
a pile of screenshotsa second memory
Your screenshots finally find their place.
In short
An AI app that reads your screenshots and turns each one into a card (a recipe, a film, a place, a product) with plain-language search and reminders.
- Problem: people save dozens of screenshots a day and never go back, so the gallery becomes a digital graveyard.
- Solution: a screenshot becomes a card of one of 13 types with its own fields, searchable by meaning in Russian and English, and turns into a task or a reminder.
- Benchmarked 19 models on 23 hand-labelled screenshots: reading got 30× cheaper and more accurate (99.1% vs 93.9%).
- 18 recorded decisions (ADRs), a PRD, market research and a pricing model with about 91% margin.
$0.0081$0.00027÷30
cost of reading one screenshot
93.9%99.1%
accuracy on 23 hand-labelled shots
13
card types, each with its own fields
Problem
A phone gallery is a digital graveyard.
A screenshot is an intention: to cook, watch, buy or visit. A week later it is lost among thousands of others. The target user is the “digital hoarder”, 25–40, active on Instagram and Telegram, with more than 5,000 screenshots in the gallery.
Competitors in the research: Pixel Screenshots, Recall, mymind, Readwise and Fabric. None of them does the one thing this audience needs: turn a screenshot into the next action.
Solution
- 01
Save
From any app through the share sheet, or in bulk from the gallery; paid plans sync the gallery automatically.
- 02
AI reads it
Each screenshot becomes a card of one of 13 types: a recipe keeps its ingredients, a film its rating, a place its address.
- 03
Find
Search by meaning in two languages (“something Italian for dinner”), full-text and vector at once.
- 04
Act
Intentions (cook, watch, read, visit, buy) become tasks with a daily digest.
Try it
Decisions
- ADR 015
Screenshot reading moved to Qwen after a 19-model benchmark
Why. Ran 19 models × 23 hand-labelled screenshots through the production prompt. qwen3.7-flash: $0.00027 per shot against $0.0081 for the previous model, and 99.1% accuracy against 93.9%. Images are priced by pixels rather than tiles, which is where the gap comes from.
Rejected: The ADR 014 choice made on two shots: it told a disaster from normal, not first place from third.
- ADR 012
The “Link” type was removed: the model never chose it
Why. Zero cards of that type out of 160 in production. “Bookmark” describes an intent, not content, and every page has a more specific type. A dead type the user can see is the worst kind of dead code.
- ADR 005
The free-tier quota is enforced server-side and atomically
Why. The client-side check had a race: five parallel saves on the last free slot all got through. A quota that costs money must live where the money is spent.
- ADR 016
A card opens on its content, not the screenshot
Why. A vertical screenshot filled the whole first screen, so the price, cooking time or concert date started on the second. People come back for the recipe, not a picture of it. The decision was checked in five interviews.
Economics
| Plan | Price | AI saves |
|---|---|---|
| Free | $0 | 5 a month + 20 at sign-up |
| Pro | $2.99 / mo | 200 a month |
| Premium | $9.99 / mo | 700 a month |
The marketing plan’s rule: paid acquisition only after D1 retention ≥ 35% and D7 ≥ 15%. The North Star is “saved and found”: the user does not just file screenshots but comes back for them.
How it is built
Expo and React Native in TypeScript; Supabase: Postgres with pgvector, pg_cron, storage and 14 edge functions. Image reading runs on Qwen, text tasks and embeddings on OpenAI, and search is hybrid. Subscriptions via RevenueCat, analytics in PostHog, errors in Sentry, end-to-end tests in Maestro, releases through EAS with over-the-air updates.
Status
Code is complete, seven in-app products are set up in App Store Connect, and the app is being prepared for submission. The snixt.com site with its waitlist is live in two languages.
Want more? Switch the reading depth to 30 min.