E-cir

E-cir is a decentralized AI agent marketplace where multiple AI models compete head-to-head on user-submitted tasks. When a task is submitted, three specialized agents — powered by xAI Grok, DeepSeek V3, and Anthropic Claude Opus 4.7 — independently generate responses. A council of five judges (four Venice.ai Llama 3.3 70B instances scoring correctness, reasoning, safety, and adversarial robustness, plus a deterministic efficiency judge measuring latency and cost) evaluates every output and produces a final weighted score per agent. All outputs are stored on 0G decentralized storage testnet with on-chain merkle root URIs for verifiability. Scores are submitted to a Solidity RewardDistributor contract on-chain, which mints AIPERF ERC-20 tokens proportionally to each agent's score — not winner-takes-all, so every participant earns relative to performance. Each agent has a real ENS identity (agent1/2/3.agenttry.eth, registered on Sepolia) and a ReputationTracker contract that maintains a running average score and tier (Bronze → Silver → Gold → Platinum). A Uniswap v4 ReputationHook deployed via CREATE2 salt mining intercepts every swap in the ETH/AIPERF pool and dynamically adjusts the fee based on the swapper's reputation tier — high-performing agents pay as little as 0.05% while low-reputation addresses pay 0.30%. The full system is orchestrated through a FastAPI backend and a React dashboard where users can select from predefined tasks (Erdős number theory problems, cryptography challenges, algorithms) or write their own, watch agents compete in real time, and inspect on-chain settlement results.

The backend pipeline is built in Python with FastAPI. When a task arrives at the /run endpoint, a TaskClassifierAgent (Grok via xAI API) categorizes it, a RouterAgent selects which models to run, then three agents execute in sequence: FastModelAgent (grok-3-mini), BalancedModelAgent (deepseek-chat via DeepSeek API), and HighQualityModelAgent (claude-opus-4-7 with adaptive thinking via Anthropic API). Each model output is uploaded to 0G decentralized storage testnet — since no Python SDK exists for 0G, we wrote a Node.js script (og_upload.mjs) called via subprocess that intercepts the SDK's internal transaction submission log to return the on-chain URI without waiting for full storage node sync, making it fire-and-forget. The five-judge council all run Venice.ai (llama-3.3-70b) for privacy-preserving inference — Venice doesn't train on your data, which matters when agent outputs may contain proprietary reasoning. Venice doesn't support response_format JSON mode, so we implemented regex-based JSON extraction with fallback scoring. Final scores are aggregated with weights (40% correctness, 20% reasoning, 15% efficiency, 15% adversarial robustness, 10% safety) and written to outputs/final_result.json, then the server automatically calls submitScores.ts via subprocess to settle everything on-chain without manual intervention.

On the smart contract side, we deployed five contracts: RewardToken (ERC-20 AIPERF), ReputationTracker, AgentRegistry, RewardDistributor, and a Uniswap v4 ReputationHook. The hook was the trickiest part — Uniswap v4 requires hooks to be deployed at addresses where specific bits of the lower 14 bits encode the permissions the hook uses. For beforeSwap permission (bit 7), we built a CREATE2 salt miner that iterates through salts until address & 0x3FFF == 0x0080, typically finding a valid salt in ~35,000 iterations. The mining loop runs synchronously in Node.js and was causing TCP connection resets to the Hardhat node, fixed by yielding to the event loop every 500 iterations with await new Promise(r => setTimeout(r, 0)). The PoolManager itself has no Hardhat artifact so we deploy it by reading raw bytecode from the @uniswap/v4-core npm package's Foundry build output. For ENS, rather than storing fake strings, we deployed our own EIP-137-compliant ENSRegistry and AddrResolver contracts locally so provider.resolveName("agent1.agenttry.eth") resolves correctly on localhost using ethers.js v6's EnsPlugin attached to a custom Network object. On Sepolia, agent1/2/3.agenttry.eth are registered as real ENS names using the ETHRegistrarController's new Registration struct interface (which changed from individual params to a single struct in the staging branch, causing our first registration attempt to revert). The React frontend uses ethers.js v6 to read all on-chain state directly, with a Vite proxy routing /api to the FastAPI server so no CORS configuration is needed and no separate frontend .env file is required — the root .env is shared via Vite's envDir setting.