yee!

• 

  yee!

Inspiration

5-Studios is a human optimization venture studio building at the intersection Web3 and AI. These technologies represent a shift in the collective consciousness to democratize business and decentralize knowledge work. I don’t know too much about the sports industry but having worked with artists like Future and Pressa, I understood that knowledge work was a talents biggest cost and applied these ideas to this hackathon.

What it does

yee! is a personal sports agent, our current features including automating expensive knowledge work to a fraction of the cost and time. yee! democratizes knowledge work for athletes by keeping NBA regulation experts at the touch of button, personalizing sports endorsement contracts, and providing state-of-art time-series forecasting with the push of a button.

How we built it

Started off with a concept of a node-based logic editor for non-technical users to quickly build out custom tools. Initially I thought about building this out of with something like Flowise when I realized Respell was one of the sponsors. When I tried their mobile application, I noticed it wasn’t optimized for mobile support so I decided I wanted to build a simple UI to interact with LLMs on respell.ai.

JUDGES LINKING BARON DAVID SUBMISSION BELOW AS WELL

Started some quick mockups for layout ideation and hopped onto Figma to start with layouts and then moved onto wireframing. After setting up initial wireframes, I noticed we could add links so tried adding a simple stock spell from Respell and testing functionality.

After getting that working, I set up a fusion of Chain-of-Thought and Prompt Tuning to create a “prompt skeleton” for Respell and add an extra document input to the LLM have semantic references of the document it needs to create.

I was getting crunched on time and was having trouble with Respell API so I used a couple existing streamlit applications to protype functionality in the Figma.

The app simulates the idea of having a personalized knowledge base so that the user essentially has a PAGI system customized to their needs.

Challenges we ran into

I didn’t even know how to open a Google Collab last year so I’m not the most technical. It was relatively easy for me to set up a front-end and back-end separately but had blocks trying to connect the two, which why I focused to using Figma. I had trouble formatting API calls and have barely used Figma before but found myself learning quick. Since it’s a Figma file, the prototype auto prompt’s a “site message” when trying out internal features and has trouble when the user is trying to exit from features external app back to the Figma file. Also in retrospect, working on a team of one for a hackathon wasn’t the most optimal idea 😂. Also design isn’t the greatest and the protype has some small component/layout errors.

Accomplishments that we're proud of

Being a non-technical team of one and building an MvP prototype in less than 48 hours. Implementing technologies such as the NBA CBA and TimeGPT, a billion parameter time-series forecasting transformer. The demo works rather well for how far I thought I was going to get. Got to work on topics such as athletic performance by improving productivity and democratizing knowledge work for athletes, especially the Warriors, has been a childhood dream.

What we learned

Learned that the enterprise knowledge work revolution is here and if you don’t walk now you can’t fly later. A large portion of the highest-income jobs in sports can be automated by the talent themselves and that chatGPT can’t make whole React project customized to your instructions just yet.

What's next for yee!

Planned on adding this for the hackathon but was short on time and technical skills; but wanted to run the whole application on serverless cloud GPU’s with frameworks like e2b and building the prototype out into a PWA or iPhone application for streamlined use. Adding a working semantic cache for user memory, improving completions with adding pydantic objects instructions, and improving results with integrating the neural search engine for LLMs by Metaphor.Systems.

Further improvements would be to a desktop webapp running on a node-based flow builder, adding a library of API/function call tools, integrating existing embedding databases, adding more transformer models, and adding auto-training based on the semantic cache.

Built With

• chatgpt
• figma
• langchain
• metaphorsystems
• pydantic
• pypdf
• respell
• streamlit
• timegpt
• vercel