Postmark

A journaling app where every day becomes a postage stamp. Offline-first, zero-cost PWA. Being built from scratch.

Folder map

Folder	What's inside
docs/	Postmark-v2-spec.md (full project spec: data model, logic, screens, build order) and How Postmark works.html (concept explainer).
design/	The 6-screen mockup board, the raw Claude Design export, and the v2 foundation mockup. Visual source of truth.
archive/	Original downloaded zips, untouched. postmark-v2-foundation.zip holds the previous app code if you ever want to look at it.
app/	The new JavaScript build (Vite + React). cd app && npm install && npm run dev.

Stack follows the spec, except plain JavaScript instead of TypeScript. Try time travel: http://localhost:5173/?now=2026-10-01T09:00.

Install on your phone

Open https://ashmita13dutta.github.io/postmark/ in Safari (iPhone) or Chrome (Android), then Share, Add to Home Screen. It works offline after the first visit.

Useful commands (run inside app/)

Command	What it does
npm run dev	Start the app locally
npm test	Run all tests
npm run lint	Check the code
npm run evaluate	Measure the built-in (keyword) mood engine on the test corpus
npm run model	Download the note reader's model (done for you before dev and build)
npm run train:head	Retrain the note reader after the labelled notes change (-- --cv to cross-check first)
npm run evaluate:model -- file.json	Score the note reader and the keyword engine on notes it has not seen
npm run assets	Redraw the app icon, launch screens and share card

The note reader (on-device model)

Postmark reads each note in two ways. The built-in way matches words (feelings.json, topics.json) and always works, offline, from the first launch. The note reader is a small language model (e5-small-v2, about 34 MB) that understands whole sentences, so "Ordinary Friday. Pizza night with the roommates." reads as content instead of bored. It runs in a web worker on the phone: your notes never leave the phone, and after the first download it works offline.

• It is opt-in. Today asks once, then downloads about 30 MB (the model plus its runtime, compressed; about 50 MB of space once unpacked), kept in the browser's cache. You can switch it off or on again under You.
• Where it is not available (not downloaded yet, still loading, an old browser) the built-in way is used, exactly as before.
• The model is not in git. npm run model downloads it from Hugging Face at one pinned revision and checks every file's hash (app/scripts/model/manifest.json); CI does this on every build.
• The model produces 384 numbers per note. A small "head" (app/src/data/feelingHead.json, 57 KB, generated, do not edit) turns them into a probability for each of the 16 feelings. Retrain it with npm run train:head after changing the labelled notes in app/tests/corpus/.
• When you tell Postmark a note felt different ("Not quite?"), it keeps the note's words and your choice on the phone. A later note that reads almost the same follows your choice. You can copy those notes from You and use them as real test notes: save them as app/tests/corpus/real.json, then npm run evaluate:model -- tests/corpus/real.json.

Measured on notes the reader had never seen (two held-out sets, 200 notes): the feeling lands in the accepted set 83% of the time against 57% for the built-in way, and the right feeling is in the reader's top three 95% of the time (79% for the built-in way). Those notes were written by the developer, not by you, so your real notes are the test that counts.

Licenses for every font, emoji set and image: app/ASSETS.md.