Inspiration
First, we try to find some problem that can be assisted and potentially solved with AI models. Also, the solution needs to be fully on-chain. Then, we start to think about how we can utilize LLM for on-chain data with context that's not too big.
We find that the price discovery process for rarer NFT is long if not listing at the right price. But we have a lot of metadata as well as trading data that are helpful to do a comparable analysis between the floor NFT and the rarer NFT to create a reference for pricing.
What it does
First, It estimates the reasonable price for rarer NFT in the collection based on comparable valuation model. Second, it creates bid on NFT marketplace automatically if it's seriously undervalued based on your intent in natural language.
How we built it
- use floor price and the metadata of the NFT that's listed as floor price as the refence point.
- take additional factors like 7D volume, floor price change, last sale price, the historical floor price for the last sale and etc. This way, AI model can construct a comparable model to evaluate how much better the target NFT than the floor NFT.
- tell the AI model the instructions to provide a reasonable price for the target NFT
- utilize the On-chain AI Oracle from Ora Protocol to get back result in our smart contract
- Parse the result and send to our intend engine so we can create a bid / listing on NFT marketplace automatically.
Challenges we ran into
- the factors that we need to consider in the comparable model
- hard to parse the right information from the result
Accomplishments that we're proud of
we have made the first fully on-chain auto trading bot based on LLM-assisted valuation for the digital assets
What we learned
We learned how to use AI inference on-chain with OAO contract and build fully on-chain application.
What's next for LLM-assisted valuation based NFT trading bot
- Fine Tune: For every valuation inference, we can update the model with the real price when the evaluated NFT is executed. AI model will learn from the actual sale price and update the weights.
- Reinforcement Learning for the bot. For every valuation inference, we can reward the model if the sale price of the evaluated NFT is actually sold at the predict price
- extend the asset class to ERC20 assets
Log in or sign up for Devpost to join the conversation.