HeptagonAI

Heptagon AI is a platform that links data providers, computing power contributors, and AI model builders, allowing for efficient and collaborative AI development. We use proof of learning scheme to guarantee verifiability of model training. The datasets get posted on Filecoin by contributers. The trainer than trains the model and computes proof of learning. Proof of learning gets verified inside Phala's TEE. Finally we use Flare Data Connector to check which users have contributed data to the dataset and how valuable the data was and reward the users accordingly.

The project uses Filecoin's lighthouse storage (we also implemented EVM compatible onramp contracts) to store datasets. We than implemented training and proof of learning for the simple image prediction AI model made in pytorch. We trained the model and calculated proof of learning. We verified proof of learning inside Phalas TEE simulator - we used tdx_quotes to sign the valid proof and verify it on the Solidity smart contract (we mocked Automata's verifier). Finally we used Flare network's FDC to verify the contributions that the users have made (quality of the data is estimated by its variance) to the model and FTSO to calculate accurate prices of the rewards to pay. We made our own attestation type for this. For the frontend we used Nextjs and React. We used Blockscout to display transactions.