On-Device Security & Architecture
How PocketAI isolates inference, preserves battery, and eliminates cloud surveillance vectors.
1. Native ARM Sandbox Execution
PocketAI bundles compiled 64-bit llama.cpp native binaries optimized with ARM NEON instructions. The native C++ library executes in an isolated process sandbox managed directly by the Android Linux kernel. No unauthenticated third-party code is ever downloaded or dynamically evaluated.
2. SHA-256 Supply-Chain Verification
Large language model weights can be gigabytes in size. To prevent corrupted downloads or man-in-the-middle tampering, every model download:
- Downloads into a temporary
.partbuffer with HTTP Range resume support. - Calculates a continuous SHA-256 cryptographic digest across the entire byte payload.
- Refuses to load and discards any file whose checksum deviates from our verified manifest.
3. Memory Safety & Context Capping
Android process death from Out-Of-Memory (OOM) killer is the primary risk in mobile LLM execution. PocketAI incorporates an active memory guard that checks available device RAM before allocation:
Total RAM Needed ≈ Model_Weights (e.g. 1.06 GB) + (Context_Length * 2 * n_layers * n_heads * head_dim * precision) + App_Overhead (~300 MB)
4. Zero Network Dependencies for Pro Entitlements
Cloud license verification requires constant pinging of authentication servers. PocketAI uses local HMAC-SHA256 signature verification. Your entitlement is verified locally on the CPU in less than 2 milliseconds with zero battery drain.