ParaMind

ParaMind is a neural network built completely on-chain. It acts like a smart contract GPU, capable of processing complex tasks such as image classification directly within a smart contract.

https://github.com/taijusanagi/2025-eth-online/blob/main/contracts/contracts/ParaMind.sol

In this hackathon, I trained an MNIST image classification model and uploaded it to the smart contract, achieving more than 90% accuracy through on-chain inference!

This project uses:

• Parallel processing with Arcology Devnet
• Real-time transaction visualization with Envio HyperIndex
• Built with Hardhat v3

If this technology can be integrated with AI agents, it means the AI agent doesn’t just operate on the blockchain — it lives in the blockchain! Theoretically, this implementation proves that it’s possible to run very complex models, like ChatGPT, entirely on-chain.

The core of the system relies on atomic transactions and parallel processing. Neural networks require a massive amount of computation — each unit in a fully connected layer interacts with every other unit, resulting in an O(n × n) computational cost.

To compute those,

// In Python, we can do this:
a @ W

// But in smart contracts, we usually have to do this:
for (uint256 i = 0; i < 2; i++) {
    for (uint256 j = 0; j < 2; j++) {
        sum += a[i] * W[i][j];
    }
}

This usually causes an out-of-gas error when the computation becomes too complex.

To solve this, each transaction focuses on a single computation. The computation itself is actually quite simple, each transaction computes just one weight × input operation.

The key component here is Arcology’s parallel processing, which allows all computations to run simultaneously while the summed values are updated in parallel. With just 5–6 requests, the entire image classification process can be completed efficiently.

All on-chain computations are logged as Events, allowing Envio HyperIndexer to aggregate and visualize the data in real time. All contract development was done with Hardhat v3.