Haoxian Yan

Every gallery app makes one of two choices: sync everything to a cloud you do not control, or defer to the system media library and inherit its organisation model — which is to say, folders and a date. Neither lets you actually annotate a picture and find it again three years later by what is in it.

ImageDB is the third option. One HTML file, one SQLite database, one folder of image files, all on your device. The deliberate minimalism is the point: an archive you cannot open in ten years is not an archive.

Ingestion — FAB add, hold-for-camera, drag & drop, paste, URL-linked images, Android share-sheet intake, and a batch import wizard with per-image and bulk field editing.

Books — multi-page collections from images, ZIP or PDF. Reader with scroll / left-to-right / right-to-left modes, per-page copy, save and share, and export as PDF or a paneled image.

Organising — albums, free-form tags with autocomplete drawn from every tag ever used, captions, live search, and filters by album, tag, type or date. A global Tag Manager renames or deletes tags library-wide.

Viewing — responsive grid with XS–XL card sizes and a time-grouped feed. Detail view is a swipeable card with a ghost preview sliding in from the side. Lightbox with scroll, pinch and double-tap zoom plus drag-to-pan.

Editing — crop canvas with undo stack and animated-GIF output, a GIF editor (speed, frame drop, quality, trim, crop), and a Panel Images combiner with per-image crop and re-grouping.

Data portability — scoped multi-part ZIP export, merge-safe multi-file import, and a verify-and-repair scan. Trash is a soft delete with 30-day auto-purge.

Search & privacy — on-device text recognition (OCR) makes words inside screenshots and receipts searchable; a Safe area behind Android's fingerprint/PIN prompt stores its images encrypted (AES-256-GCM); an optional app lock; smart albums.

Other — multi-select batch operations, an "Organise Quiz" daily tagging challenge with a Library Health score, dark/light theme, multi-language with system auto-detect, and a configurable backup nudge.

Images and thumbnails live in a layered blob store: the Capacitor filesystem on Android (app-private, invisible to gallery apps), OPFS on modern browsers, and IndexedDB as a legacy fallback. On upgrade, blobs migrate transparently from IndexedDB to OPFS with no user action.

Metadata, the hashbook and settings live in SQLite compiled to WebAssembly via sql.js. Every snapshot is written to a .tmp file, read back and checked with PRAGMA quick_check before it replaces the primary. Startup validates the primary and can recover the staged or last-known-good copy; an unrecoverable database is reported as an error rather than silently presented as a healthy empty library.

SHA-256 deduplication maps every image hash to a blob ID, so importing the same file twice reuses the existing blob.

Frontend — single HTML file, no build step. www/ is a generated artefact stamped with the version from package.json.

Native shell — Capacitor for Android with custom Java plugins for clipboard images (a content:// URI backed by FileProvider, so other apps paste real pixels), durable share-sheet intake, gallery saving, biometric privacy levels, on-device ML Kit OCR, and Google Play Billing.

Libraries — sql.js, fflate, gifenc, pdf.js, and the Capacitor filesystem, camera, share, clipboard and app plugins. All bundled; nothing is fetched from a CDN at runtime.

No account, no ads, no analytics of my own, no ImageDB-operated sync service. Originals, thumbnails and metadata live in app-private storage; Safe images and thumbnails are also encrypted at rest.

Text recognition — OCR runs on the device, but Google's ML Kit SDK sends device/app information, a per-install identifier and performance diagnostics to Google over HTTPS. Google Play services may download recognition models.

Purchases — the optional one-time Advanced Tools purchase is processed by Google Play; I never see payment details.

Android system backup — allowBackup is intentionally on. If you enable Android backup, the OS may include ImageDB's data in your backup. I do not turn that on and do not operate the destination.

Linked images — an entry saved as a URL is fetched from its source host when displayed, exposing a normal network request to that host. You can make linked entries offline from Backup & Data.

In short, data leaves the app only through OCR diagnostics, Play purchases, linked images, Android backup you enable, or a share/save/export you start.

Note apps and budgeting apps both want to be the place your life lives, and both want that place to be their server. The moment you accept that, your data's longevity is tied to someone else's business model — and personal finance records are about the last thing I want sitting in a startup's database.

Memora inverts it. The server is yours; it runs on a machine you control and writes files you can read. The phone app is a client, not the source of truth. If I stop maintaining it tomorrow, the Flask app and your files keep working.

Variants — one codebase serves three modes: memos, finance and planner. The backend routes each to its own store, selected by query param, header or referer.

Sync — an event-sourced operation log. Clients push operations and pull everything since a given op ID; server state is rebuilt purely by replaying the log, with snapshots for compaction. Because the log is variant-agnostic, a new data type needs no new sync code. Conflicts resolve last-write-wins per entry.

Storage — JSON document stores plus SQLite, currently mid-migration toward SQLite throughout. Live data sits in data/ and uploads/, both permanently excluded from version control.

Frontend — one shared source tree used by both the server and the mobile build. Includes memo capture, a finance view, OCR for receipts, and an optional AI chat that can create entries from uploaded images.

Testing — Playwright browser tests plus a backend assertion suite covering sync and cross-variant isolation.

There is no Memora cloud. I do not run a server for this, cannot see your data, and have no way to. Your notes and finances live on hardware you own, and the app talks only to the server address you type in.

Cleartext HTTP is allowed — deliberately, because most people self-host over a home LAN without a certificate. That is a real trade-off: on an untrusted network, use HTTPS or a VPN rather than plain HTTP.

AI features are opt-in and bring-your-own-key. Chat, OCR-assisted entry and image analysis call whichever provider you configure (OpenAI, Gemini or DeepSeek) using your API key. Leave the key unset and those features stay off; set one and what you submit goes to that provider under their terms, not mine. Everything else works with no key at all.

Backups are yours. Data directories never leave your machine on their own — which also means nobody else is making backups for you.

Brain-training apps have a reputation problem, mostly earned: the category is full of subscription funnels, inflated neuroscience claims, and daily-streak mechanics designed to farm engagement rather than attention.

I wanted the actual tasks — Schulte tables, stop-signal, interval recognition — without the apparatus around them. These are well-established psychological paradigms and they work fine as small self-contained games. No accounts, no streak guilt, no claims about making you smarter. Just the task, a timer, and your own previous score.

• Schulte Table — find numbers in sequence on a grid as fast as possible. Visual scanning and concentration.
• Stop Signal — respond fast to go signals, withhold on stop signals. The classic executive-control paradigm.
• Whac-A-Mole — reaction speed and target acquisition under time pressure.
• Animal Ranger — study animals with randomised profiles, then match each to its profile. Associative memory under load.
• Lotus Trail — memorise a path across lily pads that vanishes after a preview. Spatial working memory.
• Chord Scout — hear a synthesised chord, name it from three options. Major / minor / diminished discrimination.
• Interval Birds — birds sing an interval; identify it. Fully synthesised with the Web Audio API.
• Rhythm Sync — birds count in, audio cuts out, and you tap exactly on the next silent beat. Internal timing without a cue.

Nothing is collected. No accounts, no analytics, no ads, no network calls. The games hold no personal data — there is nothing to transmit and no server to transmit it to.

Scores and theme preference stay in local storage on your device. Audio is synthesised in-browser, so even the sounds require no downloads. Uninstalling removes everything.

This started as a sandbox: I wanted to work out the interaction design for calendar-plus-todos-plus-projects without committing to a data model or a sync story. Building against local storage first meant I could throw away layouts cheaply.

It worked, so it is now being ported into Memora as a third sync variant alongside memos and finance — each event, todo, project and habit becoming one entry in the same event-sourced operation log. The standalone prototype stays around as its own thing for now.

Local storage only. The prototype has no backend at all: everything you enter stays in your browser and never leaves the device. Clearing site data erases it — there is no copy anywhere else.

Once the port into Memora lands, it inherits Memora's model: your own self-hosted server, your own data directory.

I wanted to find a specific panel — the one with a particular expression, or a particular gadget — in a collection far too large to page through. Filename search is useless for this; the information you want is inside the image.

It turned into a nice demonstration that a full OCR-plus-embedding search pipeline fits comfortably on one desktop machine. Nothing here needs a datacentre.

1. Panel extraction — every page is read from its ZIP archive, flood-fill/contour analysis detects individual panels, then PaddleOCR extracts the Chinese text into SQLite.

2. Tag enrichment — each untagged panel goes to DeepSeek Chat for 5–15 candidate Chinese semantic tags. Each candidate is deduplicated against a global vocabulary using BAAI/bge-m3 cosine similarity at a 0.90 threshold: close matches reuse the canonical tag, new ones extend the vocabulary.

3. Embedding build — every panel is encoded into a 1024-dim float32 vector with BAAI/bge-m3, preferring canonical tags over raw OCR text. A one-time offline step, roughly an hour on CPU.

On first request the server loads every embedding into an in-memory numpy matrix and caches the tag vocabulary. The unified search endpoint combines three modes, each contributing a normalised 0–1 score: 词 (keywords), always on; 意 (vector), cosine similarity on L2-normalised embeddings, on by default; and AI (semantic), where DeepSeek maps the query onto vocabulary tags, off by default.

The combined score is the sum of active methods normalised by how many are active, and negative exclusions (-term) work across all three. Result cards show extracted text, tag chips with matches highlighted, and a score badge.

Runs entirely on your own machine against your own files. The one exception is the optional AI tagging and AI search mode, which send text (never images) to DeepSeek using your own API key. Keyword and vector search work fully offline once the index is built — the embedding model runs locally.

A zero-dependency embeddable reader for images and PDFs, originally extracted from ImageDB so every project would stop maintaining its own copy. Scroll, left-to-right and right-to-left page modes; double-tap and pinch zoom; keyboard controls; long-press for a page preview with copy/save/share; configurable theme, background colour and page separator. One JS file and one CSS file, no bundler, with pdf.js as the only optional dependency.

2,129 MIT-licensed SVG icons across three collections (Scarlab Oval Line, Solar Broken Line, Coolicons Line Oval), validated as well-formed and shared across the apps. A handful were repaired or redrawn in-set where the upstream source had gone missing. Icons are inlined into each app's stylesheet at build time rather than fetched at runtime, so nothing phones home for a glyph.