StakeSocial

(Sorry my flights is in a couple of hours so I had to leave early and couldn't stay for judging, but hopefully you enjoy! Had some issues with wifi at the airport so couldn't get a video in 4 minutes without speeding it up but if you can test it out yourself on Farcaster!)

We all love prediction markets, but these are largely focused on global affairs and sports. Me and my friends often spend a lot of time having informal bets with each other about things like sports (intramural games), if two friends will get in a relationship with each other, if a friend will fail a class, a round of poker and other fun things! We thought it would be fun to build something out to facilitate this, and make it easy to do! The entire app revolves around group chats with friends where any chat can have as many markets as they can. The main issues there might be that we wanted to address with this

• Resolution: How do we do verification in a fair way. It can't be simple majority since in that case if the majority lose they could hijack a bet, and it obviously can't be a single person. We thought the most fair way would be requiring unanimous verificaiton. To make it simpler users can upload pictures of their friend's all giving a thumbs-up (resolve yes) or thumbs-down (resolve no) and using the facial embeddings, figure out the resolution. Additionally the agentic feature allows you to upload a picture of proof with anti-watermarking to reduce doctored pictures. But really, if you're friends, hopefully you don't cheat!
• Bias: A person could bias the result if they knew about the bet. As a result when making a prediction you can choose to 'hide' the market from certain people in the chat who it's about.
• What if everyone loses: we had the idea of just counting it as a 'house' stake but thought that was a bit too cheeky, so we ended up deciding that it goes into a 'dinner fund'

The base app was made through a Celo-templated version of Farcaster, which made it much easier to work with and deploy the app! We used XMTP for all the chat-messaging between the friends, and managing friend groups and a little bit with agents. We used facial landmarking (YOLO v7) and embedding (EigenFace) for the computer-vision based verification and image extraction. The agentic workflow was built with simple tool calls through OpenAI (gpt 5 with minimal). We used Pyth for the nondeterminism in the casino and degen modes for a little bit more fun, and it let it be in a way where the user can confirm there's no foul play happening with how the randomness is done! Used hardhat for testing the contracts before deployment and estimating fees and so on.