Confidential Computing Architecture
Sightline integrates deeply with NEAR AI Cloud’s Confidential Computing infrastructure. Rather than treating AI as a black box or relying on corporate trust promises, Sightline uses hardware-isolated CPU memory encryption to guarantee mathematical privacy for sensitive document processing.
NEAR AI Cloud vs. Standard Commercial Cloud AI
A side-by-side comparison of privacy boundaries, hardware isolation, and custody guarantees.
1. Execution Host & Isolation
Inference executes inside CPU-encrypted Virtual Machines (Intel Trust Domain Extensions). Host kernel, hypervisor, and adjacent tenants cannot inspect guest memory.
Inference executes on general shared virtual machines without memory encryption. Host hypervisors and system processes have complete read access to memory buffers.
2. Operator & Third-Party Access
Neither NEAR AI infrastructure engineers, datacenter technicians, nor root administrators can view plaintext documents or model activations during execution.
Cloud provider administrators, internal platform engineers, and human QA / moderation review teams can inspect inputs, chat logs, and uploaded files.
3. Data Retention & Storage Lifetime
Enclave RAM is completely ephemeral. Nothing is persisted to databases, object storage, or server disk. Data is purged when the session terminates or when the user clicks Forget.
Prompts and uploads are routinely retained in API logging databases, diagnostic caches, backup volumes, and training datasets for future model fine-tuning.
4. Verification Proof
Client browser verifies Intel TDX quote before sending documents and validates model-generated Ed25519 cryptographic signatures on completion. Verifiable math, not promises.
Privacy relies entirely on corporate legal policies and reputation. End-users have zero cryptographic means to independently audit or verify data handling.
3-Stage Verifiable Execution Pipeline
Client-Side Sanitization & Nonce Generation
When taking photographs on mobile devices, EXIF metadata, camera device tags, and GPS coordinates are permanently purged in browser memory via HTML5 Canvas before network transit. A fresh, cryptographically random 32-byte session nonce is generated to prevent replay attacks.
Hardware Enclave Inference (NEAR AI Cloud)
Requests are dispatched non-streamed to TEE-hosted vision and language models (e.g. Qwen/Qwen3.8-27B and z-ai/glm-5.3-flash) running inside Intel TDX enclaves. CPU-level encryption guarantees memory isolation from the host OS and hypervisor.
In-Browser Cryptographic Verification
Before displaying results, the browser independently checks the hardware attestation quote (8 client-side checks including TDX authenticity, non-debug attributes, and nonce binding) and verifies the enclave's Ed25519 signature over the completion bytes.