Loops HouseSealroom

NEAR HACKATHON
Links
Team
1 member- ASOwner
Aswin Prem Kumar
Overview
Sealroom helps organizations compare sensitive records and use AI without handing over datasets or giving agents unrestricted access. It combines private matching, confidential AI analysis and verifiable access controls on NEAR.
The problem appears across industries: banks checking overlapping exposures, consulting firms managing competing clients, businesses reviewing supplier relationships, and research teams comparing confidential records. These are potential applications of the same approach; our prototype demonstrates it through conflict checks between law firms.
In that example, a partner moves firms. Both firms need to identify client conflicts without exchanging their client lists. Sealroom finds exact matches through private set intersection and uses a model inside a NEAR AI Cloud trusted execution environment to identify less obvious relationships. Inputs are encrypted to the attested model key. Inside the enclave, the model reads the records and returns entry numbers rather than names.
The NEAR contract verifies both organizations' signatures and up to 48 inference signatures per check. Proof bundles let reviewers reconstruct requests, verify responses and compare the derived verdict with its recorded hash. Attestation checks cover Intel TDX, runtime measurement replay and NVIDIA GPU evidence.
After matching, access rules keep restricted documents out of the AI agent's context. Decisions form a tamper-evident log sealed on NEAR, with scoped agent keys and signed user authorization.
The prototype runs on NEAR testnet with a public demo. Broader deployments would need industry-specific policies and validation. Document access enforcement currently trusts the application server; the conflict-matching inputs follow a separate encrypted path to the model.