Idle models, listed on-chain
A consumer GPU hosting open weights is enough to list, with no permission needed. The on-chain order book is persistent and states model, unit price, capacity and proof tier.
Years of building exchange wallets, market-making and risk-control systems. For the past two years I've focused on where AI meets Crypto: agents that wager, battle and settle on their own, and AI-assisted market analysis. My current focus is LLMChain: pooling idle LLM capacity on-chain for agents, with a verifiable receipt for every call.
There are plenty of idle GPUs and self-hosted open-weight models out there, and more and more agents firing off thousands of inference calls a day. LLMChain is an independent L1 that connects the two: anyone with a model can list it on-chain; agents pick a provider by model, price and proof tier and pay it directly; every call produces a Verifiable Service Receipt (VSR), and the receipt is the settlement. Humans trade on invoices. Agents trade on VSRs.
A consumer GPU hosting open weights is enough to list, with no permission needed. The on-chain order book is persistent and states model, unit price, capacity and proof tier.
An agent reads the order book, opens an x402 payment channel, and the provider accepts and executes off-chain on the spot. No waiting for blocks; only settlement goes on-chain.
Every completed call produces a VSR, the only object that ever goes on-chain. Funds settle against receipts, with no platform custody.
Covert challenges, tiered fidelity proofs and stake slashing make swapping a small model in behind a flagship listing a losing trade.
The foundation only builds a thin verification-and-settlement base layer. Revenue splits, routing and the rest of the ecosystem evolve in contracts, driven by the community. Whitepaper v1.1 is just out, with plenty still marked TBD. If you have idle GPUs, or opinions on verifiable inference, payment channels, TEEs or zkML, come file issues, poke holes, and help build it.
A sports prediction engine built on multi-agent debate: seven AI analysts with different cognitive biases argue over three rounds, combined with a Zep knowledge graph and Polymarket on-chain smart money, to predict moneyline, spread and totals.
github.com/hellojun/PolySport →An open battle protocol for AI agents: any agent can register, wager, fight and settle by reading a single Skill file, paying over HTTP via x402. The protocol never custodies keys. Game developers who submit new game types earn a lasting share of fees.
github.com/hellojun/bout.network →Recognizes the match while you browse football markets on Polymarket and gives multi-expert debate analysis in a sidebar. A second line lets you borrow USDC against Polymarket positions, with AI-suggested LTV and auto-close before kickoff.
github.com/hellojun/kpax-ball →A cluster of Go market-making bots (quoting, hedging, arbitrage, self-trading, risk, alerts), plus the integration of a Python multi-factor backtesting and execution system. A team project; I own the overall design and integration.
Team project · not publicWrote my first public project, usdt_wallet, and sent a PR to the EOS token list. I was already interested in how money moves safely on-chain.
Built WeChat Native Pay and a smart photo-studio system for a photography brand.
Worked on the 1Box NFT platform, the Xoasis NFT marketplace (including an Ethereum wallet SDK for Unreal Engine) and the gamelink chain-game project.
Started ex-wallet: a complete wallet backend in Go, with deposits and withdrawals, a signing service, a risk service, MPC and KMS, BTC/ETH/BSC/TRON/XRP node integration and an ops console. I wrote the website and part of the admin API myself.
Built gchain, a game app chain on the Cosmos SDK, and won first place at the hackathon of the Cosmos HK Accelerator's 2nd Chinese developer training. Also started GreatGainGo, building oracle contracts and nodes.
Wrote inscription tools for Celestia, IoTeX and Cosmos, and started CelestLoan and an inscription marketplace. Also wrote a small tool that polls Binance listing announcements every second and goes long on OKX, which unexpectedly became my most-starred repo.
Started memes-ai, a microservice multi-chain meme-coin trading platform (ETH/BSC/Base/Solana/Tron, with MEV protection). Also built Solana node services, a TradeGPT-style trading assistant, and two Texas Hold'em AIs (one on LangChain, one on the Claude API).
PolySport, the Bout protocol, KPAX Ball, and market-making and multi-factor quant systems. In October, published the LLMChain whitepaper: pooling idle LLM capacity on-chain for agents.
An independent L1 that pools idle LLM capacity on-chain for agents, producing a Verifiable Service Receipt (VSR) for every call and settling natively.
View repository →Multi-agent debate × knowledge graph × on-chain smart money: a sports prediction engine.
View repository →Open Agent Gaming Protocol. A Hono + Next.js + Drizzle monorepo; the contract handles USDC escrow and settlement.
View repository →Reference implementation of an external game server for the Bout protocol: implement IGame, wrap it in HTTP, register with the protocol.
View repository →Polymarket football AI-analysis extension with a position-collateralized lending contract (Foundry, UUPS upgradeable).
View repository →Polls Binance announcements every second, spots new perpetual listings, and places long orders on OKX.
View repository →A Cosmos SDK game app chain; first place at the CHA 2nd Chinese developer training hackathon.
View repository →A USDT wallet from 2019. Where it all started.
View repository →Repository count · personal and organizations
Code on GitHub, me on Telegram, or just email.