AI ZK-Compute

This project implements a decentralized hyper-parameter search where off-chain workers train a tiny neural net, produce a Groth16 proof of their reported accuracy, and the on-chain AiOrchestrator verifies proofs and pays rewards.

How it works

1. Client opens a job with a grid of hyper-parameters (e.g., 3 learning-rates × steps) and funds a bounty.
2. Workers each claim exactly one grid slot, train locally, generate a witness + Groth16 proof, and submit their accuracy + proof to a contract on-chain.
3. Orchestrator contract verifies proofs, records accuracies, and when closed, pays 50% to the best accuracy and splits the rest equally among other proven participants.

Components

The project codebase includes:

• a Solidity orchestrator,
• a Circom circuit + verifier,
• a Python worker,
• and a simple client scripts to open requests.

How to run the project

Prerequisites

• Modern Linux or Mac OS X operating system
• Rust 1.90+
• Node.js 18+
• pnpm 8+
• Python 3.9+ (venv recommended)
• circom 2.1.x (install via Cargo - see below)
• snarkjs 0.6.x (see below for installation instructions)
• ts-node 10.x (see below for installation instructions)
• ptau file for zk

circom (via cargo -> First Install Rust):

• cargo install --locked --git https://github.com/iden3/circom

snarkjs:

• pnpm i -g snarkjs

ts-node:

• pnpm install -g ts-node

Make sure pnpm and .cargo/bin directories are in your executable PATH.

Execute

• Open terminal window, navigate to the project directory and execute
  • make clean to reset your environment to the initial state (if necessary)
  • make deps to prepare the environment
  • make startup to compile contracts and launch blockhain nodes locally.
• Open another terminal window in the project directory and execute make run to compile off-chain workers, deploy contract onto the (local) blockchain, and submit sample ML training request.

Makefile targets

• make startup # compile blockchain side, start blockchain nodes
• make run # deploy contract on the node, submit sample request
• make clean # remove all compiler artefacts and environment configuration
• make mlp-zk # compile ZK-proofs module
• make nde # compile contracts and start hardhat node
• make depl # deploy to localhost
• make req # open a sample request (grid + bounty)
• make workers-XX REQ_ID=X # compile and run XX workers, submit requests with ID X (number). Makefile defines 3, 6, 9, 12 worker targets. E.g. make workers-12 REQ_ID=0

Project layout

• contracts/
  • AiOrchestrator.sol # assigns tasks, verifies proofs, pays rewards
  • AccVerifier.sol # auto-generated by snarkjs (do not edit)
• circuits/
  • MlpHoldoutAcc_256.circom #
  • MlpHoldoutAcc_256_js/ # wasm + witness generator (generated)
  • MlpHoldoutAcc_256_final.zkey # proving key (generated)
• client/
  • open_request.ts # opens a request with a hyper-parameter grid
• node/
  • compute_node.py # worker: claim, train, prove, submit
• Makefile
• .env