Phulax
Your defi Guardian agent that never sleeps. Protect. Detect. Withdraw.
Demo
About
This project is meant to safeguard users in yield protocol. I myself have been victim of defi hacks on various yield protocols and so I wanted to build something that can help that. What I built is a guardian that watches over every transaction that hit the pool you are in. You deposit through `PhulaxAccount, a smart contract where the agent key can call exactly one function withdraw(adapter)` and funds are hard-coded to return to you. This is fine as worst case, the agent gets compromised and forces an exit. A Keeperhub worflow has most of the steps, it fetches the latest transactions, runs it through a detection pipeine which checks for invariants, oracle deviation, vector similarity against historical exploits, and a classifier we fine-tuned. If risk crosses the threshold, the keeperhub workflow is set to withdraw the funds and append a log to 0g. The guardian is also an iNFT so it can be monetized and transferrable. Here is the demo: - https://46.225.141.153/web/what-is-it -> explains the project - https://46.225.141.153/ -> a fork of keeperhub with 0g integration and my custom workflow. To Sign in using username: patrick@bytalabs.com and password: password. You will find 2 workflows, one is to show the 0g nodes crearted, the other one "Phulax Guardian" is the one I'm talking about - https://46.225.141.153/web -> frontend to see what the agent is doing
How it's made
I combined the power and extensibility of Keeperhub. First I added 0g as a first class citizen workflow into Keeperhub (which I will make a PR for) so we can do inference/kv get and puts and append logs to 0g. This was useful for my use case. Tech used: Keeperhub, nodejs, 0g cli and docs Then I created a library of past defi attacks, stored those on 0g storage, used 0g to fine tune a model on those attacks (using Qwen 2.5). Because 0g doesn't allow yet to run inference on the fine tuned model I had to spin up a server for that inference step, otherwise this would be on-chain (or on network). Tech used: Python, FastAPI, uvicorn, transformers, PEFT, PyTorch, Pydantic v2 Then I created the smart contracts that were needed for this, a fake lending pool, the Phulax smart contract and a fake token with scripts to send transactions to them. Also created the iNFT contract. Tech used: Solidity, foundry, iNFT Then I wired everything together using Keeperhub which was very easy as it's simply drag and drop. I had to fine tune some steps to ensure that the agent doesn't by default withdraw, it was a lot of fine tuning on the different detection tiers. But I've always thought it's better to be safe than sorry (forego a few cents of APY to keep your principle!) I created a nextjs frontend to visualize it all, along with SSE events so we can understand what the whole flow is doing.