Loops HouseRAMX

NEAR HACKATHON
Links
Team
1 member- SAOwner
Sarnav Kansal
Overview
Problem: big AI models don't fit in one laptop's memory, while nearby laptops and GPUs sit idle. Pooling them needs more than networking: identity, fair billing, payments that AI agents can make on their own, and limits on what a compromised agent can spend.
What we built: RAMX pools the RAM and GPUs of nearby Apple and NVIDIA machines into GPU Pools. Each laptop holds a slice of the model's layers, so the pool runs a model none of its members could load alone. People and agents use it through an OpenAI-compatible API, paying with credits or per request in USDC via x402. Providers earn USDC for the answers their pool serves.
How NEAR fits in (all live on testnet, every step linked on NearBlocks in the README):
- Payments: agents pay per request in USDC via x402; meta-transactions through our relayer cover gas.
- Guarded agent wallets: an agent holds only a function-call key that can call pay on our escrow contract. Budget, per-call cap, allow list and pause are enforced by the contract; a prompt-injected agent that tries to pay a stranger or overspend is refused on chain.
- Audit trail: every balance change is a signed event written to our ledger contract; anyone can verify a tx with one command and no keys.
- Payouts: revenue is split by rule and posted as a Merkle root with USDC in one ft_transfer_call; providers claim their share themselves.
- Private AI: private prompts run on NEAR AI Cloud in a TEE with fresh-nonce attestation, optional end-to-end encryption, and only salted hashes on chain. A TEE judge spots-checks pool answers.
- Login with any NEAR wallet (NEP-413).
- AI Swarms: 25 agents with their own NEAR sub-accounts bid for compute in a sealed-bid market, with LLM-chosen bids and signed commit-reveal.