Loops HouseSigstack: Verifiable Private AI for Signal

NEAR HACKATHON
Links
Team
1 member- BAOwner
bagelface
Overview
Sigstack is a Signal bot for voice transcription and group translation. It runs in an Intel TDX enclave on Phala Cloud, so your messages are decrypted only inside hardware nobody can inspect, not even us. The problem is that a TEE is only useful if users can tell which code is in it. For the hackathon we made that checkable by anyone, from a Signal message down to the GPU.
What we built:
-
A Shade Agent contract on NEAR testnet (agent.sigstack.testnet). On boot, the bot generates a key inside the TEE and registers it with a TDX quote. The contract verifies the quote against Intel's collateral, and accepts it only if the build's compose hash and the CPU are on the owner's approved lists. Every approved build and every registration is public on-chain.
-
Model verification. Inside its own TEE, the bot verifies NEAR AI Cloud's models: their TDX quotes, and NVIDIA's verdict on their GPUs. It checks that every translation is signed by an attested model instance.
-
!verify. Send the bot
!verify. It replies with a fresh TDX quote over your text, its compose hash and the model evidence. You can check the quote on proof.t16z.com and the rest on our verifier page. -
A verifier page and build archive. A static page reads the contract live. For every approved build, it hashes the published compose file in your browser and shows GitHub-signed provenance for each image. CI re-checks that every build traces to a commit in this repo.
Known limits: NEAR AI doesn't sign Whisper transcripts yet, NVIDIA's verdict is trusted over TLS, one owner key approves builds, and images aren't bit-for-bit reproducible.
Sigstack predates the hackathon; only the NEAR work above is submitted (diff linked in the README).