BlinkCode AI

Optical file transfer as an offline-first PWA. A file is compressed into a compact "File DNA" container, painted as SpectraCode colour matrices, and recovered by pointing a camera at the sender's screen — no server, no network after the first launch. Small payloads fit in a single static frame; larger ones loop as a numbered frame sequence the receiver collects while scanning.

file ─▶ order-1 arithmetic coder ─▶ File DNA ─▶ frames ─▶ RS(255,223) ─▶ colour grid
                                                                            │
              file ◀─ arithmetic decoder ◀─ CRC-32 ◀─ frame collector ◀─ RS ◀┘
                                                          ▲
              camera frames ─▶ bullseye detection ─▶ homography ─▶ colour correction

Run it

npm run build:model   # (re)train and package the downloadable model bundle
npm run serve         # http://localhost:8080
npm test              # codec + full optical loopback tests

Open the app, pick a file (or paste text), hit Compress, then Show SpectraCode on the sender and Start scanning on the receiver — the receiver decodes continuously and shows how many frames it still needs. Without a second device, the loopback self-test button pushes every rendered frame back through the real optical decoder.

What is actually implemented

Piece	Status
One-time model download, Cache Storage + IndexedDB, fully offline afterwards	real
Order-1 context model + arithmetic coder (lossless, any file type)	real
Perceptual image mode: 64-value latent, 8-bit quantised	real (linear DCT autoencoder, see below)
SpectraCode render: 36-colour palette, bullseye markers, calibration strip	real
Reed-Solomon RS(255,223), corrects 16 bytes per block	real
Marker detection → homography → 3×4 colour correction → cell classification	real
Multi-frame streaming: chunk header, out-of-order collection, CRC-32 verify	real
Two grid densities (120×80 / 60×40) signalled by a density patch and auto-detected	real
Continuous camera auto-scan, wake lock while displaying, Web Share of results	real
Workers for model loading, compression, decompression and optical decode	real
Installable PWA with precached shell	real

Honest limits

These matter more than the pitch, so they are stated plainly in the app's "How it works" tab too:

• One SpectraCode frame holds 5,111 payload bytes (1,097 in robust density): 120×80 cells × 5 bits = 6,000 bytes raw, minus Reed-Solomon parity and the 18-byte frame header. Bigger files are split across frames, so throughput is roughly 5 KB × frame rate — about 15 KB/s at 3 frames per second, and every missed frame costs a full loop.
• 1000:1 lossless compression of arbitrary files is impossible. Lossless coding cannot go below the entropy of the data; a universal 1000:1 compressor would let you recurse to one byte. Real measured ratios here: ~2–4× on text, JSON and source code, ~1× (slight expansion) on already-compressed data such as JPEG, ZIP or MP4.
• Extreme ratios only exist in perceptual mode, which is genuinely lossy: an image becomes a 64-value latent and comes back as a soft, 32×32-detail reconstruction. That is a 1000:1+ ratio and a very different picture.
• The "neural" model is a mock with a real interface. The bundle format, one-time download, offline caching and the coder's model API are production shaped; the weights inside are trained order-1 byte statistics plus DCT quantiser scales (~129 KB), not a 50 MB neural net. Order1Model in src/lib/arith.js and the transforms in src/lib/image-codec.js are the two swap points for a real learned predictor / autoencoder — nothing else changes. TensorFlow.js is intentionally not bundled: no runtime CDN dependency is allowed, and shipping a placeholder network would add megabytes without improving compression.
• Camera decoding is validated in simulation, not against physical phone optics: the test suite renders a code, applies perspective tilt, a colour cast and vignetting, and decodes it end to end. Real-world glare, motion blur and low-end camera sensors are what the robust density exists for; they may still need per-cell voting and multi-frame averaging.

Layout

src/
  index.html app.js styles.css sw.js manifest.webmanifest
  lib/
    arith.js          arithmetic coder + order-1 model (swap point for a neural predictor)
    codec.js          File DNA encode/decode orchestration
    format.js         container header
    image-codec.js    perceptual latent encoder/decoder
    geometry.js       linear solver, homography, colour-correction fit
    palette.js        the 36 reference colours
    rs.js             Reed-Solomon GF(256)
    spectracode.js    layout, symbol packing, renderer
    spectra-decode.js marker detection → payload bytes
    storage.js        one-time model download + offline persistence
    stream.js         frame header, CRC-32, out-of-order frame collector
    surface.js        canvas / buffer drawing targets
  workers/            model loader, compression, decompression, optical decode
  models/ucg-v1.bin   downloadable model bundle
tools/                build-model.js, serve.js
tests/                codec + end-to-end optical loopback

Licence

BlinkCode AI is source-available under the BlinkCode AI Personal Use License, not an open-source licence:

• Personal, academic and non-profit use is free, including modifying and sharing the code.
• Attribution is required — keep the licence file and credit Dr Sohil Momin (@drsamonline) with a link back to this repository wherever you credit authors.
• Commercial or enterprise use needs written permission from the author; ask via https://github.com/drsamonline for a commercial licence.