← Back to Projects
leptons1618@anishgiri.dev:~/projects/comic-sudoku cat project.toml
Comic Sudoku
Point your camera at a Sudoku puzzle and it solves it on the spot — a TensorFlow.js CNN that reads the grid in the browser, then ships as an Android app and a PWA.
- category
- Computer Vision
- year
- 2023
- status
- Stable
- stack
- 8 tools
- highlights
- 6
01Highlights
- A CNN reads the nine-by-nine grid from a camera photo entirely in the browser — no image is ever uploaded.
- Three training runs are committed under `training/runs/`, including the Keras weights, the TF.js export, and a `metrics.txt` per run.
- The current run scores 97.6% validation accuracy and is benchmarked against rotation, dim light, bright light and noise — noise is the weak spot at 89.2%.
- Weights are SHA-256 verified between the trainer and a fresh subprocess before export, so a corrupted artefact cannot ship silently.
- Capacitor packages the same build as a native Android app, with the Gradle project and Play Store release docs in the repository.
- Apache-2.0, and it installs as a PWA too — one codebase covering desktop, mobile and the app store.
02Built with
- 01TypeScript
- 02React
- 03Vite
- 04TensorFlow.js
- 05OpenCV.js
- 06Capacitor
- 07Tailwind CSS
- 08Keras