Loops HouseSecureLambda

NEAR HACKATHON
Links
Team
1 member- JAOwner
Jason Stanley Yoman
Overview
Problem: AI teams cannot test a model on private data without exposing either the data or the model. Sellers will not hand over datasets, buyers will not hand over models, and AI agents have no safe way to pay per use.
What we built: SecureLambda, a pay-per-run evaluation service. A seller lists a dataset and a buyer (often an AI agent) uploads a model. Both are encrypted on the client with the evaluator's public key, so the API and the storage never see plaintext. The agent asks to run the model, receives an x402 price quote (HTTP 402), signs the payment, and gets back only summary scores (accuracy, AUC, F1) and a receipt. Failed runs are never charged.
How NEAR fits in: NEAR is the payment and identity layer, running live on testnet. The agent signs a NEP-366 delegate action (a meta transaction) that calls ft_transfer on wNEAR (NEP-141, wrap.testnet). A relayer submits it only after the score is ready and pays the gas, so the agent needs no gas and is charged only for results. We found that a native Transfer inside a delegate is charged to the relayer, so we pay in wNEAR to make the buyer's own balance fund the run. Sellers register via NEP-145 storage deposit. Login is a NEP-413 signed message with a full-access key, with no passwords or API keys. A real paid run is verified end to end: the seller received exactly the price and the settlement transaction is public.