Loops HouseKuvos

NEAR HACKATHON
Links
Team
1 member- CHOwner
Chinmoy Kalita
Overview
Problem: after a hospital stay, families get dense discharge instructions. AI can turn them into a daily routine, but in healthcare two questions block it: who saw the document, and who decided what the plan is? An AI draft must never silently become someone's care plan.
What I built: Kuvos on NEAR. A patient uploads a discharge PDF or a photo of a prescription. Confidential TEE models on NEAR AI Cloud read it. The SDK verifies the Intel TDX + NVIDIA attestation, encrypts the request end to end to the attested model, and checks the model's provider_tee signature on every reply. A reply without that signature fails the job. Every extracted item is shown beside its highlighted source text, and nothing becomes a task until a person confirms it.
How NEAR fits in: confirming creates a salted commitment to the plan version, its evidence and the TEE signatures. The user can sign that digest with their NEAR wallet (NEP-413 via NEAR Connect, no gas). Daily check-ins and every decision by the Tomo assistant are committed too. Tomo runs a published rule gate first, and its rules hash is stored on-chain. Commitments are batched into Merkle roots and anchored to a Rust contract (kuvos-near.testnet) using a function-call-only key. Anyone can verify a proof in the app or offline. No health data goes on-chain.
Live: https://app.kuvosai.com