Demokratia

Current DAOs suffer from extreme voter apathy - a driving force behind DAO centralisation. At Demokratia, we address voter apathy through automation, specifically through making personalised, privacy-preserving AI agents in DAO voting easy to implement.

Our belief is that through automation we can include more voices in the decision-making process, many of which have not been heard before. In the long run, automation is also a means to make democracies better at decision making and competitive in the face of autocratic challenges.

Here’s a high-level overview how Demokratia with a hypothetical DAO:

For each DAO voter, we generate a feature-rich, privacy-preserving representation locally on their device and send it to a trust-minimised zkML coordinator service. The coordinator takes into account the pending poll context and then verifiably votes on the voter’s behalf while reflecting the voter's individual preferences specified through the representation.

There are three components to our architecture:

1. On-device:

A simple IOS app that serves as the entry point for DAO voting. The app also contains a voter registration form and the BERT encoder model.

Upon logging in with WalletConnect, the voter can choose a DAO and fill out a custom voter registration form to specify their ideological preferences. Following that, BERT generates an embedding locally.

The voter then send the embedding to our trust-minimised coordinator, who automates voter decision making. If the voter is not satisfied with the automatic vote, they can always override it. In other words, the voter can vote in a DAO (in the demo we used MACI contracts) through our IOS app manually.

1. Blockchain:

We use Arbitrum to store the verifier contracts for delegator proofs and custom MACI contracts.

1. Server:

We run a modified MACI coordinator service. Our server-side zkML model accepts embeddings from user devices, collect the relevant poll context, and outputs user-encrypted autogenerated votes into the MACI poll contract.

During tallying, the coordinator proves that they have correctly computed the total result as the correct combination of original user messages and the valid autogenerated votes.