The first challenge that we tackled was the scoping of the overall product and ecosystem. Our initial idea was to enable agent-to-agent transactions and the resale of agents. We soon realized that it would be too complex to implement, especially for the hackathon. We decided to focus on LLMs rather than agents for the POC because models do not rely on external APIs and therefore do not require resetting API keys upon ownership transfer.

Another challenge was the trade-off between decentralization and mitigating the risks of rug pulls. We decided that the platform mandates deploying and serving inferences from fine-tuned models on a shared cloud infrastructure. This centralized approach ensures that the MLOps of fine-tuned models won't be disrupted in the event of IP ownership transfers. Having a centralized GPU cloud infrastructure poses the risk of having a single point of failure. This can be mitigated by leveraging a decentralized computing network for deployment and serving inferences.

We create ERC-721 tokens to govern the monetization and resale of LLMs. We turned the ERC-721 token into ERC-6551 and leveraged the Unlock Protocol to issue time-bound tokens from the token-bound account (TBA).