connecting...
Your challenge
Draw this word โ†“
โ€”
Draw on canvas.
Click blurred emoji to reveal hint.
Time remaining
30
seconds
remaining
Rounds
Score
0/0
rounds played
0
Model thinks it's
โ€”
start drawing...
waiting for strokes...
Prediction history
Kalakari
A sketch recognition experiment.
Draw. The neural network guesses.
Can you communicate with AI?
Built by Shiven Saini
Powered by SketchNet
shiven.dev
50 categories ยท draw any of these
๐Ÿฑcat
๐Ÿถdog
๐Ÿš—car
๐Ÿ house
๐ŸŒณtree
๐ŸŸfish
๐Ÿฆbird
โœˆ๏ธairplane
๐Ÿšฒbicycle
๐Ÿ•clock
โ˜€๏ธsun
๐ŸŒ™moon
โญstar
๐ŸŒธflower
๐ŸŽapple
๐ŸŒbanana
๐Ÿ•pizza
๐ŸŽธguitar
๐ŸŽฉhat
๐Ÿ‘Ÿshoe
๐Ÿ˜elephant
๐Ÿฆ’giraffe
๐Ÿงpenguin
๐Ÿฌdolphin
๐Ÿฆ‹butterfly
๐Ÿ“strawberry
๐Ÿpineapple
๐Ÿ‰watermelon
๐Ÿ‡grapes
๐Ÿ“ทcamera
โ˜Ž๏ธtelephone
๐Ÿ’ปlaptop
๐Ÿ“บtelevision
๐Ÿ›‹๏ธcouch
๐Ÿช‘chair
๐Ÿ›๏ธbed
๐Ÿšชdoor
๐Ÿ–ผ๏ธpicture frame
๐Ÿชœladder
๐ŸŒ‰bridge
โ›ตsailboat
๐ŸšŒbus
๐Ÿš‚train
๐Ÿšhelicopter
๐ŸŽˆhot air balloon
โš”๏ธsword
๐Ÿ‘‘crown
๐Ÿ’Ždiamond
โณhourglass
๐Ÿ•ฏ๏ธcandle
5
Rounds
30s
Per Round
96%+
Model Acc.
ONNX
Runtime
Tip: draw fast and simple shapes for best guesses.
Round complete
โ€”
Final score
โ€”
Under the hood
SketchNet
~2.1M
Parameters
64ร—64
Input Size
96%+
Val Accuracy
50
Classes
Forward pass โ€” Input โ†’ Output
Input
Sketch rendered
from strokes
(1,64,64)
โ†’
Block 1
Conv 3ร—3
ResBlock
MaxPool /2
(32,32,32)
โ†’
Block 2
Conv 3ร—3
ResBlock ร—3
MaxPool /2
(64,16,16)
โ†’
Multi-Scale
4 parallel branches
Inception-style
Concat โ†’ 128ch
(128,16,16)
โ†’
Block 3
DW-Sep Conv
ResBlock
MaxPool /2
(192,8,8)
โ†’
Block 4
DW-Sep ร—2
ResBlock
MaxPool /2
(384,4,4)
โ†’
GAP
Global Average
Pooling
(384,)
โ†’
Output
Linear + Softmax
50 classes
(50,)
Multi-scale block โ€” stroke-width invariance
1ร—1 Conv
Fine details. Stroke tips & sharp corners.
3ร—3 Conv
Medium strokes. Most common width.
5ร—5 Conv
Thick & gestural. Bold drawing styles.
Pool + 1ร—1
Context features. Surrounding structure.
All 4 branches run in parallel then concatenate โ†’ 128 channels. A clock drawn thick or thin looks the same to the network.
Design decisions
Residual blocks
Skip connections prevent gradient vanishing and regularize โ€” critical for sketch data where everyone draws differently.
DW-Sep convolutions
Blocks 3 & 4 use depthwise separable convolutions โ€” ~9ร— fewer params than standard Conv2d with similar accuracy.
Global avg pool
GAP gives translation invariance โ€” a cat drawn in any corner of the canvas produces the same feature vector.
Mixup training
Training blends pairs of drawings (ฮฑ=0.4). Forces smooth boundaries between confusable pairs like clock/sun and cat/dog.
SGD + OneCycleLR
SGD with Nesterov momentum + OneCycleLR warmup/cosine decay. Converges in ~25 epochs from random init.
Client inference
Model exported to ONNX and runs entirely in-browser via ONNX Runtime Web (WebGPU / WASM). No server needed.
Author
Shiven Saini
Kalakari Project ยท 2026