I did it.

I can flag scam tokens on Base.

(With 94% accuracy)

Building, not sure what yet but sharing my journey.

Prev. @rainbowdotme, @metamask

Santiago, Santiago Metropolitan Region, Chile

Instead of attempting smart contract audits, I developed a neighborhood-based algorithm, based by the principle:

If you interact with a scammer, you're also likely a scammer.

I trained graph neural network models to understand:

• How tokens interact

• The behavior of their neighborhoods

• Subgraph patterns that indicate scams

Here’s one example of the scam behavior my model can flag:

1. A “Token Deployer” wallet gets funded by a “Malicious Wallet” to deploy the “Target Token.”

2. Adds 100 ETH liquidity to a decentralized exchange.

3. Removes liquidity for 118 ETH, sending it back to the funding wallet.

The algorithm also detects subgraphs of related tokens interacting via middle-man wallets, flagging broader networks of malicious behavior. You can see 4 other tokens in red.

🚩 And finally, turns out the "Malicious Wallet" has thousands of transactions following this model.