Loops HouseBlindsight

NEAR HACKATHON
Links
Team
1 member- PROwner
Prosper Chimaobi Akam
Overview
BlindSight is a confidential data marketplace on NEAR. Publishers list sensitive datasets in sealed vaults, such as financial results, medical cohorts and security audits, and set a price per question. Buyers, whether people or AI agents, pay in NEAR to ask a question and get an AI answer. They never see or download the raw records.
The problem: valuable data can't be shared safely. Handing over a file means losing control of it. BlindSight sells answers instead of files.
How it works:
The buyer approves two native NEAR testnet transfers in one wallet step: the vault price to the publisher and a small platform fee. The server checks both on chain (signer, receiver, amount, status, age) before any AI runs. Inference runs on NEAR AI Cloud (DeepSeek V3.2) inside an Intel TDX enclave with GPU attestation. BlindSight does not take this on trust. On every question it checks the gateway and model attestations against Intel's certificate chain, using a fresh one-time code. It encrypts the question with Oblivious HTTP to a key signed by the attested gateway, because standard HTTPS ends outside the enclave. It then verifies the enclave's signature over the exact question and answer. The buyer gets a "Verified private answer" receipt showing each check. Publishers upload datasets that are encrypted at rest (AES-256-GCM). Only the publisher's wallet can remove a vault, proven with a NEP-413 signature. NEAR in use: Wallet Selector, near-api-js, FastNear RPC, NEAR AI Cloud private inference and its attestation and signature APIs.
Honest limit: BlindSight's own server still decrypts the dataset to build the prompt. Moving that step into a TEE is next.