With AI bringing automation with LLM chatbots, trading agents, etc., in Web3, we desperately need a verifiability layer for these AI agents. Think of a scenario where AI agents can get compromised, and then they can fire random on-chain interactions. To solve this, while inferencing, the AI agent will use our ZKML toolkit to create a proof along with the inference. This ZK proof is used to verify the computation done by the AI agent to generate the inference. These proofs are first verified on-chain and then the AI agent can perform the desired action on-chain.

To build this toolkit, first I created a small ZKVM with all the required computations. Currently the ZKVM supports dense layers used in deep learning models. Using this ZKVM, I built a transpiler that can convert the ML computations into a ZKVM trace, which is used in our Plonky3 circuits to make a ZK STARK proof. We have also used Mersenne31 and CirclePCS for optimised ZKML proving.