Scribbl
Onchain coprocessor AI to judge your doodling skills
Created on 15th February 2025
•
Scribbl
Onchain coprocessor AI to judge your doodling skills
The problem Scribbl solves
AI systems are often plagued by bias, centralization, and opaque decision-making processes that can lead to unfair outcomes. Our dApp tackles these issues by running inference on a fully verifiable Linux machine, ensuring that every judgment is made in an unbiased and decentralized manner.
Using our onchain coprocessor AI, we provide a level playing field for doodlers—where the contest is fair, competition is transparent, and every creative stroke is evaluated solely on merit. Our system transforms the traditional, secretive AI evaluation process into an open, verifiable, and fun experience. So whether your doodle is a masterpiece or a merry mess, our decentralized judge is here to cheer you on and reward every quirky creation!
Challenges we ran into
We encountered hurdles while training the model and running it inside the Cartesi Machine. We aimed to train our model on a popularly known Google's open-source Quick Draw dataset. The tutorials around it though were heavily focused on having a full TensorFlow environment. Unfortunately, taking this route is not the best practice we could do inside our target RISC-V machine. After some research, we ended up finding a TensorFlow Lite pre-trained model on the same dataset which was running great for Android apps. This lightweight runtime worked like a charm for us and gave pretty good results.
Technologies used
