Snap a photo. The AI solves it.
Take a picture of any sudoku — printed in a book or hand-drawn on paper. The AI reads every cell, then either solves the whole puzzle or whispers the next certain move into your ear.
Upload a photo, get the answer
Recognised grid
Click any cell to correct an OCR mistake before solving.
Three steps from photo to answer
1. Computer Vision finds the grid
Adaptive thresholding + morphological line extraction isolate the printed grid. A perspective warp maps the four outer corners onto a 540×540 square; a polynomial dewarp then removes the residual page-curl typical of book photos.
2. CNN reads every cell
The warped grid is sliced into 81 cells. A small CNN trained on MNIST plus synthetic printed digits in 17 fonts classifies each cell. Cells below 85 % confidence are flagged for review.
3. Solver fills the blanks
For the full solution, a backtracking solver with MRV heuristic finds the answer in under 20 ms on a typical puzzle. For hints, a logical solver returns the next naked-single or hidden-single move with a human-readable explanation.
Want to run it yourself?
The full pipeline — OCR, CNN, solver — is open-source. Clone it, train the CNN on your own photos, or bolt the API onto your own app.
📦 GitHub repository