AI-powered Token Engineering

"AI-powered Token Engineering" is a project to use LLMs in the context of token engineering.

Token Engineering is the design, verification and optimization of token-based systems. Here, modelling and simulations are important to build systems with rigor, stress test our systems, understand if the system properties hold in various scenarios, find the optimal parameter settings, and more.

For our hackathon project "Ai-powered Token Engineering", we use LLMs to provide a natural language interface to token models and support the simulations process using custom tools.

Our main focus is to get from a fuzzy human input to a prompt that triggers planning and execution agents, which is based on typical data science methods and Python functions. The idea if these agents orchestrating the simulations process is inspired by BabyAGI and "Plan-and-Solve" paper. For the skills and reasoning process we're using the semantic kernels skill tree approach. Our agents

• have access to system parameters and documentation of an Python token supply model (cadCAD https://cadcad.org/)
• can run scenarios and propose potential changes to system parameters to achieve a certain metric target value

Our hackathon achievements:

• combine Open AI GPT 3.5 and Google flan t5 large model using the hugging face API
• enable agents to answer on questions about the model using benchmark data, accessible as data asset on Ocean Protocol

Built With

• cadcad
• gpt3.5
• langchain
• openai
• python
• radcad