I'm Jay Wang, a safety researcher at OpenAI. I design, code, research, and write. I received my PhD from Georgia Tech in 2024.
My work focuses on making AI more accessible, interpretable, and accountable, by designing and developing novel Interactive Interfaces for people to easily and enjoyably interact with Machine Learning systems at scale.
AgentVault is what happens when you let AI agents spend real money but keep your hand on a kill switch. It's a treasury protocol on Solana: fund one USDC vault, register each agent with a per-task budget cap, and every payment an agent makes gets checked against that cap on-chain before it clears — no private keys handed to autonomous code, no surprise drains. Trip the kill switch on any agent and its spending access is revoked instantly, while its full history stays on the books. Every payment leaves a permanent on-chain receipt you can audit on Solana Explorer. Prefer to drive it from code? A TypeScript SDK and CLI run the whole flow headless.
Under the hood it's a small, sharp stack. An Anchor program in Rust holds the vault and enforces the budgets; vaults, agents, payments, and receipts all live as Program Derived Addresses, so the rules belong to the chain, not a server. A Next.js dashboard reads everything live from devnet — create a vault, fund it, register an agent, or hit the kill switch straight from your wallet. It's the kind of project I like most: real on-chain guardrails for AI that are genuinely useful and a little fun to run.
Follow along: https://github.com/xiaohkk/agentvault
I love open source. I lead and maintain the following open-source research
projects on GitHub. 
⬇️
Please reach out if you have any questions or want to chat!
|
|
| ||||||||||||||||||||||||||||||||||||
|
|
| ||||||||||||||||||||||||||||||||||||
|
|
| ||||||||||||||||||||||||||||||||||||
|
|
| ||||||||||||||||||||||||||||||||||||
|
|
| ||||||||||||||||||||||||||||||||||||
